Documente Academic
Documente Profesional
Documente Cultură
EXTENDED ABSTRACTS
EUROCAST 2005
Edited by
Alexis Quesada-Arencibia
Roberto Moreno-Díaz jr. and
José Carlos Rodríguez-Rodríguez
© 2005 IUCTC Universidad de Las Palmas de Gran Canaria
ISBN: 84-689-0432-5
REGISTRO: 05/1325
Preface
The concept of CAST as Computer Aided Systems Theory, was introduced by F.
Pichler in the late 80's to encompass those computer theoretical and practical
developments as tools to attack problems in System Science. It was thought as the
third component (the other two being CAD and CAM) that will provide for a
complete picture of the path from Computer and Systems Sciences to practical
developments in Science and Engineering.
The University of Linz organized the first CAST workshop in April 1988, which
demonstrated the acceptance of the concepts by the scientific and technical
community. Next, the University of Las Palmas de Gran Canaria joined the
University of Linz to organize the first international meeting on CAST, (Las Palmas
February 1989), under the name EUROCAST'89 and proved to be a very successful
gathering of systems theorists, computer scientists and engineers from most of
European countries, North America and Japan.
EUROCAST 2005 continue with the approach tested last Conferences (2001 and
2003). There are different specialized Sections which, in this occasion, are devoted to
Formal Approaches in Modelling (C1), Tools and Models for Social Systems (C2),
Intelligent Information Systems (C3), Systems Applications Components (C4),
Information Applications Components (C5), Cryptography and Spectral Analysis
(C6), Computer Vision (C7), Biocomputing (C8), Intelligent Vehicular Systems (C9),
Robotic Soccer (C10) and Robotics and Control (C11).
The event has been possible thanks to the efforts of the chairmen of the Workshops in
the selection and organization of all the material. The organizers must express their
acknowledgement to all contributors and participants, to the invited speakers for their
readiness to collaborate, to the sponsors and in sum, to all those which will contribute
to the success of the Conference as a pleasant and useful scientific, technical and
human event.
Table of Contents
Degenerate Arrays.........................................................................................7
M. Miró-Julià
i
One formal systemic approach in software design. Case study:
Extreme programming.................................................................................29
C. Halbich
ii
Meteorological Image Descriptors ..............................................................58
J. L. Crespo, P. Bernardos, M.E. Zorrilla, E. Mora
iii
An Open Low-Cost Parallel Architecture for Unsupervised
Learning in Information Retrieval.............................................................100
R Mones, N Vázquez, I. Díaz, J. Ranilla, E.F. Combarro, E.
Montañés
iv
V. Information Applications Components
v
Remarks on Calculation of Autocorrelation on Finite Dyadic
Groups by Local Transformations of Decision Diagrams .........................169
R. Stankovic, M. Karpovsky
vi
Automatic Braille Character Recognition: BrailLector .............................201
N. Falcón, C.M. Travieso, J.B. Alonso, M.A. Ferrer
VIII. Biocomputing
vii
Convergence of Iterations..........................................................................235
P. Cull
viii
Fast Road Sign Detection using Hough Transform for Assisted
Driving of Road Vehicles..........................................................................269
M. A. García, M. A. Sotelo, E. Martín-Gorostiza
X. Robotic Soccer
Advances in Robotics................................................................................272
P. Kopacek
ix
A Homeostatic-Adaptive Approach for Controlling Robotic
Systems .....................................................................................................301
J.D. Hernández-Sosa, J. Lorenzo, A.C. Domínguez-Brito, M.
Castrillón-Santana
Author Index............................................................................................327
x
!
"
#
$
%
$
!
" # $% & ' !
(
)
% *
+,,-.
/
#
012
34
56
7 2
*6
8
9
: 0
;2 82
:
5
+. 56
* 82
< =(=> 92*
(?
(?
+/@@=.
=
#
'&
4
2 1 8
4
2 (
6) /@@> 9
+/@@>.< A(=
> 4 '2 34 B: )
%
+,C.< (/@
-
'B
?
)
%
+,C/.
D
E5 '2 F
%
2 B
:
< 2* 2 B
:
F
(
%
)
(
2
A +,A-.< ==(==/
A ?
& '9
!
*) 2 2 E
"(
* 5
2 2 E
4 GH +,A,.
C
#
' 8 2 2
?
4
2 *6
2 *) 9 2 :2
8
8*6
* : / +,CA.< C,(C,-
Approximation Problems Categories
2
Computing Institute
UFRGS – Porto Alegre, Brazil
{blauth, laira}@inf.ufrgs.br
http://www.inf.ufrgs.br/
Extended Abstract
Acknowledgments
This work was partially supported by: PUCRS – Pontificia Universidade Catolica
do Rio Grande do Sul, Brazil, UFRGS – Universidade Federal do Rio Grande
do Sul, Brazil and CNPq – Conselho Nacional de Desenvolvimento Cientifico e
Tecnologico, Brazil.
References
1. G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela and M.
Protasi, “ Complexity and Approximation – Combinatorial Optimization Problems
and their Approximability Properties”, Springer-Verlag, 1999.
2. J-M. Cordier and T. Porter. Shape Theory: Categorical Approximations Methods.
Ellis Horwood Ltd, 1990.
3. M. R. Garey and D. S. Johnson, “Computers and Intractability - A guide to the
Theory of NP-Completeness”, Bell Laboratories Murray Hill, New Jersey, 1979.
4. D. S. Johnson, “Approximation Algorithms for Combinatorial Problems”, J. Com-
put. System Sci. 9, 1974. p.256-278.
5. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Categorias dos Prob-
lemas de Otimização. In: XIV SBES - Workshop on Formal Methods - WFM2000,
3. Oct. 04-06, João Pessoa, Paraı́ba. Proceedings... Brazil: SBC, 2000. p.104 – 109.
6. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Optimization Prob-
lems Categories. In: Formal Methods and Tools for Computer Science - EURO-
CAST’2001. Las Palmas de Gran Canaria, Canary Islands, Spain, R.Moreno-Diáz
and A.Quesada-Arencibia Eds., 2001. p.93 – 96,
7. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Optimization
Problems Categories. EUROCAST’01, LNCS 2178. Springer-Verlag, R.Moreno-
Diáz, B. Buchberger and J-L. Freire Eds., 2001. p.285 – 299.
8. L. A. S. Leal; D. M. Claudio; P. B. Menezes and L. V. Toscani. Modelling the
Approximation Hierarchy to Optimisation Problems Through Category Theory.
International Journal of Computing Anticipatory Systems, v. 11, Belgium, 2002.
p.336 – 349.
9. C. Rattray. Identification and Recognition Through Shape in Complex Systems.
LNCS 1030, Springer-Verlag, 1995. p.19-29.
10. C. Rattray. Abstract Modelling Complex Systems. Advances in Computing Sci-
ence, Systems: Theory and Practice. R. Albrecht, Ed., 1998. pp. 1-12.
Computation of Partial Automata through Span
Composition?
Karina Girardi Roggia, Marnes Augusto Hoff, and Paulo Blauth Menezes
1 Extended Abstract
References
1. Robinson, E., Rosolini, G.: Categories of partial maps. Information and Computa-
tion 79 (1988) 95–130
2. Asperti, A., Longo, G.: Categories, Types, and Structures - An Introduction to
Category Theory for the Working Computer Scientist. MIT Press, Cambridge,
USA (1991)
3. Peterson, J.L.: Petri Net Theory and the Modelling of Systems. Prentice-Hall,
Englewoods Cliffs, New Jersey (1981)
4. Meseguer, J., Montanari, U.: Petri nets are monoids. Information and Computation
88 (1990) 105–155
5. Menezes, P.B.: Diagonal compositionality of partial petri nets. In: 2nd US-Brazil
Joint Workshops on the Formal Foundations of Software Systems, Eletronic Notes
in Theoretical Computer Science v.14 (1998)
6. Hopcroft, J.E.: Introduction to automata theory, languages and computation.
Addison-Wesley (1979)
7. Adamek, J., Trnkova, V.: Automata and Algebras in Categories. 1 edn. Kluwer,
Dordrecht (1990)
8. Bénabou, J.: Introduction to bicategories. In: Reports of the Midwest Category
Seminar. Number 47 in Springer Lecture Notes in Mathematics. Springer-Verlag
(1967) 1–77
9. Hoff, M.A., Roggia, K.G., Menezes, P.B.: Composition of Transformations: A Frame-
work for Systems with Dynamic Topology. International Journal of Computing
Anticipatory Systems 14 (2004) 259–270
Neural network sensitivity analysis applied for
the estimation of sensor matrix redundancy
Introduction
Experimental
Conclusions
It was shown that the neural network sensitivity analysis may be attractive tool
for the optimization of sensor systems. It may be considered in the context of
the sensing elements selection and their working conditions setup (e.g. operation
temperature). Similar approach may be also used in the other fields requiring
some estimation of the quality of the series of information sources.
References
1. Moore S.W., Gardner J.W., Hines E.L., Gopel W., Weimar U.: A modified multilayer
perceptron model for gas mixture analysis, Sensors and Actuators B 44: 517-520,
1997.
2. Huyberechts G., Szecówka P.M., Roggen J., Licznerski B.W.: Simultaneous quan-
tification of carbon monoxide and methane in humid air using a sensor array and
artificial neural network, Sensors and Actuators B 45: 123-130, 1997.
3. Di Natale C., Davide F.A.M., D’Amico A., et al.: A composed neural network for
the recognition of gas mixtures, Sensors and Actuators B 24-25: 808-812, 1995.
4. Jervis B.W., Desfieux J., Jimenez J., et al.: Quantification of gas concentrations in
mixtures of known gases using an array of different tin-oxide sensors, IEE P-Sci.
Meas. Tech. 150 (3): 97-106, May 2003.
5. Zou X.B., Zhao J.W., Wu S.Y., et al.: Vinegar classification based on feature ex-
traction and selection from tin oxide gas sensor array data, Sensors 3 (4): 101-109,
April 2003.
6. Martin M.A., Santos J.P., Agapito J.A.: Application of artificial neural networks to
calculate the partial gas concentrations in a mixture, Sensors and Actuators B 77:
468-471, 2001.
7. Sobański T., Saffarpour N.: Analysis of gas mixture using sensor array and artificial
neural networks. In: Cybernetics and systems 2002. Proceedings of the Sixteenth
European Meeting on Cybernetics and Systems Research. Vol. 1.: 97-100, Vienna
2-5 April 2002.
8. Żurada J.M., Malinowski A., Usui S.: Perturbation method for deleting redundant
inputs of perceptron networks, Neurocomputing 14: 177-193, 1997.
9. Figaro Gas Sensors, TGS 800 series specification, Figaro Engineering Inc. 1997.
10. Roussel S., Forsberg G., Steinmetz V., P. Grenier P., Bellon-Maurel V.: Optimi-
sation of electronic Nose Measurements. Part I: Methodology of Output Feature
Selection, Journal of Food Engineering 37: 207-222, 1998.
11. Szczurek A., Szecówka P.M., Licznerski B.W.: Application of sensor array and
neural networks for quantification of organic solvents vapours in air, Sensors and
Actuators B 58: 427-432, 1999.
Fuzzy Modeling for Coal Seams
A Case Study for a Hard-Coal Mine
Luis Argüelles, Jose M. Rivas, Javier Toraño, Carlos González, R. García, Teresa
de Pedro, J. Antonio Martín
Abstract: This work presents a set of linguistic variables for modeling some
geologic and morphological features of a coal seam. Since self-advancing
winning coal systems and their interactions with coal faces represent extremely
complex situations where traditional mathematical models are unable to offer
working solutions, we explore the viability of using Fuzzy-logic based
techniques in order to obtain a good-enough model that allows technicians at
mine to get a better understanding of the variables at play and to foresee the
coal production.
Introduction
A research project 1 has been awarded in 2004 to Carbonar S. A. by the “Centro para
el Desarrollo Tecnológico Industrial” (CDTI) agency of the Spanish Ministry of
Industry for financially supporting the research. The goals of the project are to
analyze, model and simulate the longwall exploitation system merging finite element
As exposed, one of the goals of the project is to simulate the behavior of the longwall
equipment with fuzzy logic. Fuzzy logic is a technique whose power to perform
control of systems where the model is not well known has already been proven. In
this work we attempt to use fuzzy logic not to control the longwall, but to model its
interplay with the existing geomorphologic conditions. In order to do this, the project
is planned in three phases:
1. Simulation of the coal winning system. There are many variables influencing
the coal winning activities. This part of the project will attempt to identify
them and assign linguistic values to them. Several output variables will be
established as, for instance, the number of shearer passes per labor shift, or
the daily advance in longitude for the coal seam resulting from winning the
coal. After that, it will be necessary to establish the rules linking the input
and output variables.
2. If problems are detected in the simulation by the fuzzy module under
definite conditions, a detailed physical analysis of the shield mechanism will
be run for these conditions by means of finite elements modelling
techniques.
3. A virtual reality system will show the results of the simulation.
References
1. “Fuzzy Logic in Geology”, Eds. Rober V. Demicco, George J. Klir, Elsevier Academic
Press, 2004 , ISBN 0-12-415146-9.
Data Analysis of a Distributed Solar Collector
Field
No Institute Given
Extended abstract
The purpose of this work is to analyze methods, and algorithms to control the
ACUREX solar collectors field of the Platforma Solar de Almerı́a - PSA. The
analysis of me6thods and algorithms to control the PSA solar field is a starting
point to formulate a model of solar collectors field control optimization. The sole
criterion of our optimization policy is to maximize the energy accumulated in the
oil storage tank but the optimal strategy should fulfill some constrains resulting
from physical and technical restrictions. It results in a nonlinear dynamic opti-
mization problem with nonlinear constraints, which is considered in literature
as hard. The data concerned was related to performance of various control algo-
rithms observed under changing environmental conditions. The main concepts
of the approach discussed are:
1. Prior to data analysis tasks, data preprocessing stage takes place, the pur-
pose of which is to:
– integrate data from different sources e.g., combine data from different
monitoring devicessystems into one consistent data model,
– ensure good quality of data for analysis (preliminary analysis of data
to assess data quality: discover data integrity problems, ’outliers’, miss-
ing values, etc.; design and implement data transformations to resolve
discovered quality problems).
References
1. ACUREX Field D.A.S. and Control System, Documentation version 2001,Platforma
Solar de Almerı́a, Almerı́a, Spain 2001.
2. Blanco-Muriel M., Alarcón-Padilla D.C., López-Moratalla T., Lara-Coira M.: Com-
puting the solar vector, Solar Energy 70(5):431-441, 2001.
3. Camacho E.F., Rubio F.R., Hughes M.: Self-tuning control of a solar power plant
with a distributed collector Field, IEEE Control Systems Magazine, 72-78, 1992.
4. Camacho E.F., Berenguel M., Rubio F.R.: Advanced control of solar plants, Ad-
vances in Industrial Control, Springer-Verlag, 1997.
5. Carmona R.: Analisis, modelado y control de un campo de colectores solares dis-
tribuidos con sistema de seguimiento en un eje, PhD thesis, Universidad de Sevilla,
1985.
6. Dorf R., Bishop R.: A modern control systems, Addison-Wesley Longman Inc.,
Menlo Park CA, 1998.
7. Farkas I., Vajk I.: Experiments with internal model-based controller for Acurex
Field, IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain,
2002.
8. Johansen T., Storaa C.: An internal controller for distributed solar collector fields,
IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain, 2002.
9. Pereira C., Dourado A.: Application of a neuro-fuzzy network with support vec-
tor learning to a solar power plant, IHP Program, Proceedings of the 2nd Users
Workshop, Almerı́a, Spain, 2002.
10. Silva R., Rato M., Lemos J.: Observer based time wrapped control of distrib-
uted collector solar fields, IHP Program, Proceedings of the 2nd Users Workshop,
Almerı́a, Spain, 2002.
11. Smolec W.: Fototermiczna konwersja energii sÃlonecznej, Wydawnictwo Naukowe
PWN, Warszawa, 2000.
12. Solar thermal electricity generation, Lectures from the summer school at the P.S.A.
13-17 July 1998, Almerı́a, Spain.
13. Valenzuela L., Balsa P.: A series and parallel feedfoward control schemes to regulate
the operation of a solar collector field, Serie Ponencias, Proceedings of the 1st Users
Workshop, Almerı́a, Spain, 1997.
14. Vaz F., Oliviera R.: Neves de Silva R.: PID control of a solar plant with gain
interpolation. Serie Ponencias, Proceedings of the 1st Users Workshop, Almerı́a,
Spain, 1997.
15. Zarza E. at all: The DISS Project: Direct Steam Generation in Parabolic Trough
Systems. Operation and Maintenance Experience and Update on Project Status.
Journal of Solar Energy Enginnering Transactions of the ASME, 124(2):126- 133,
2002.
Non-linear Formal Modelling
of Amphibian Tectal Cells
Extended Abstract
Lettvin, et al. were the first to find correlations between the form and function
of each of these cells [1].
Their description of the tectum was focused only to the response of tectal
cells in a prey-catching behavioral situation.
In Pedro Ramon y Cajal’s drawing of the frog’s tectum there are four main
features that appear clearly. First, the cell bodies lie in the deeper layers . Sec-
ond, the neurons seem to have a restricted dendritic tree. Third, axons leave
the tectum branching from a main ascending dendrite. Fourth, interconnections
among cells lying at the same depth is strongly suggested.
Fig. 1. (a) Schematic diagram of frog’s tectum. (b) Diagram of tectal cell interaction
processes
– a. Newness Cell
– 1. Its receptive field is approximately 30◦ in diameter, with considerable
overlap.
– 2. It yields small responses to sharp changes in illumination.
– 3. Its response frequency increases if the motion of an object is irregular.
Its response also depends on the velocity and size of the object and on the
direction of motion.
– 4. It has no enduring response.
– 5. Its adaptation time is less than 1 second.
– 6. Adaptation is erased with a step to darkness.
– b. Sameness Cell
– 1. Its receptive field is almost all the visual field, but includes a null region.
– 2. It does not respond to changes in illumination.
– 3. It responds to the presence of an object in its field by generating a low
frequency pulse train. Response is maximum to objects about 3◦ in diameter.
– 4. It ”notices” with a ”burst” of pulses all motion discontinuities. It dis-
criminates among objects, fixing its ”attention” on the one with the most
irregular movement.
References
1. Lettvin, J.Y. et al: What the Frog’s Eye Tells the Frog’s Brain. Proc. I.R.E., 47
(1959) 1940–1951
2. Ewert, J.P.: Neuroethology of Releasing Mechanism: Prey-Catching in Toads. Be-
havioral and Brain Sciences, 10 (1987) 337–405
3. Ewert, J.P.: The Neural Basis of Visually Guided Behavior. Scientific American,
230(3) (1974) 34–42
4. Lettvin, J.Y. et al: Two remarks on the Visual System of the Frog. in Sensory
Communication. W. A. Rosemblith ed. MIT Press, (1961) 757–776
5. Moreno-Dı́az, R., de Blasio, G.: Systems Methods in Visual Modelling. Systems
Analysis Modelling Simulation 43(9) (2003) 1159–1171
6. Moreno-Dı́az, R.: Conceptual Model of the Frog’s Tectum. Sensory, Decision and
Control Systems. R-548. MIT, MA. USA, (1967)
Optimization of a Class of Uncertain Systems Based on
Uncertain Variables
Z. Bubnicki
Extended Abstract
The idea of uncertain variables has been developed for a class of uncertain systems
described by a relational knowledge representation [1], [2], [3]. This paper presents a
new result in this area: an optimization problem formulated by using the uncertain
variables. The problem formulation is described after a short presentation of the
uncertain variables and a basic decision problem.
In the definition of the uncertain variable x we consider two soft properties:
~
" x ≅ x" ( x is approximately equal to x) and " x ∈ D " ( x approximately belongs to
D, i.e. its approximate value belongs to D) where D ⊆ X and X is a set of values
(vector space). The uncertain variable is defined by a set X and a certainty distribution
h ( x) = v ( x ≅ x) given by an expert, where v ∈ [0,1] denotes the certainty index of the
~
soft property, v ( x ∈ D) = max h ( x) . In the version called C-uncertain variable we
x∈ D
use the certainty index
~ 1
vc ( x ∈ D) = [max h ( x ) + 1 − max h ( x)].
2 x∈ D x∈ X − D
Let us consider a static plant with the input vector u ∈ U and the output vector
y ∈ Y , described by a relation R (u , y; x) ⊂ U × Y where the vector of unknown
parameters x ∈ X is assumed to be a value of an uncertain variable described by the
certainty distribution h (x) given by an expert. If the relation R is not a function then
the value u determines a set of possible outputs D y (u; x) = { y ∈ Y : (u , y ) ∈ R} . For
the requirement y ∈ D y ⊂ Y given by a user, we can formulate the following
decision problem: For the given R (u , y; x ) , h (x ) and D y one should find the
decision u * maximizing the certainty index that the set of possible outputs
approximately belongs to D y (i.e. belongs to D y for an approximate value of x ).
Then
~
u * = arg max v [ D y (u; x ) ⊆ D y ] = arg max max h ( x) (1)
u ∈U u ∈U x∈ Dx (u )
where D x (u ) = {x ∈ X : D y (u; x ) ⊆ D y } . The considerations for C-uncertain variable
are analogous, with v c instead of v .
The purpose of this paper is to present a formulation and solution of a specific
optimization problem for the uncertain plant under consideration. The main idea is as
follows: Consider the plant described by R(u, y; x) where one-dimensional y
denotes a quality index which should be minimized. In this case, the requirement may
be presented in the form y ≤ α or D y = (−∞, α ] where α is given by a user and
References
1. Bubnicki, Z.: Uncertain Variables and Their Application to Decision Making. IEEE Trans.
on SMC, Part A: Systems and Humans 6 (2001) 587–596
2. Bubnicki, Z.: Uncertain Logics, Variables and Systems. Springer-Verlag, Berlin London
New York (2002)
3. Bubnicki, Z.: Analysis and Decision Making in Uncertain Systems. Springer-Verlag, Berlin
London New York (2004)
4. Bubnicki, Z.: Stability and Stabilization of Discrete Systems with Random and Uncertain
Parameters. Proceedings of the 15th IFAC World Congress, Vol. E “Robust control”.
Pergamon, Oxford (2003) 193–198
5. Bubnicki, Z.: Application of Uncertain Variables in a Class of Control Systems with
Uncertain and Random Parameters. Proceedings of the European Control Conference,
Cambridge, UK (2003)
6. Bubnicki, Z.: Application of Uncertain Variables to Decision Making in a Class of
Distributed Computer Systems. In: Musen, M., Neumann, B., and Studer, R. (eds.):
Proceedings of 17th IFIP World Computer Congress. Intelligent Information Processing.
Kluwer Academic Publishers, Norwell MA (2002) 261–264
Computational Simulation of Categorical
Constructions?
Computing Institute
UFRGS - Porto Alegre, Brazil
{rodrigor,blauth}@inf.ufrgs.br
http://www.inf.ufrgs.br/
1 Extended Abstract1
References
[1] Saunders Mac Lane. Categories for the Working Mathematician. Springer Verlag,
New York, USA, 2nd edition, 1998.
[2] R. Rosebrugh and J. Bradbury. Category Theory Database Tools. Mount Allison
University, Canada, 1998. Available at http://cs.mta.ca/research/rosebrugh/ctdt/.
[3] Michael Barr and Charles Wells. Category Theory for Computing Science. Prentice
Hall International Series in Computer Science. Prentice Hall, 2nd edition, 1995.
[4] M. Fleming, R. Gunther, R. Rosebrugh. A Database of Cat-
egories. Mount Allison University, Canada, 1995. Available at
http://mathcs.mta.ca/research/rosebrugh/dbc/.
[5] Jean Piaget. Morphisms and Categories - Comparing and Transforming. Lawrence
Erlbaum Associates, Hillsdale, New Jersey, 1992.
[6] Andrea Asperti and Giuseppe Longo. Categories, Types, and Structures - An Intro-
duction to Category Theory for the Working Computer Scientist. MIT Press, Cam-
bridge, USA, 1991.
[7] Peter J. Freyd and Andre Scedrov. Categories, Allegories. North Holland Publishing
Company, 1990.
[8] D. E. Rydeheard and R. M. Burstall. Computational Category Theory Project.
Electronic edition, 1988. Available at http://www.cs.man.ac.uk/ david/categories/.
Composing Transitions Into Transactions in
UML Diagrams
Extended Abstract
The Unified Modeling Language (UML) [1–3] may be used to describe both
the structure and behavior of object-oriented systems using a combination of
notations. For the modeling of the dynamic behavior, a number of different
models are offered such as interaction, state and activity diagrams.
Activity diagrams are one of the means for describing behavior of systems
within the UML language and has been chosen here for its capability in describ-
ing flows of activities of a desired system. Broadly speaking, states of an activity
diagram mostly represent the execution of sequential and concurrent steps in a
computational process. In UML, those steps may be an action or a subactiv-
ity state. Action states represent an atomic computation without substructure
and that cannot be interrupted. Subactivity states, on the other hand, repre-
sent non-atomic execution and can be decomposed and interrupted by external
events.
When modeling concurrent or parallel systems with such diagrams, we must
be aware that basic activities of each system may be constituted by smaller
activities, i.e. transitions are conceptually refined into transactions. Nevertheless,
the UML seems to lack compositional constructs for defining atomic activities.
In this work, we concentrate on describing groups of sequential or concurrent
activities that are responsible for performing a computation, and we address
the issue of modeling transactions. We remark that in our setting the term
“transaction” denotes a certain activity of the system that might be composed by
many, possibly concurrent, subactivities. Moreover, we require this composition
of activities to be considered atomic.
The importance of discussing composition mechanisms is based on the fact,
as pointed in [4], that in order to master the complexity of systems, models
demand a form of structural decomposition. When building large systems, we
need constructs to describe systems as compositions of other smaller systems. In
this sense, when modeling a computational process, we need means of composing
subactivities both in a non atomic or atomic way.
References
1. Fowler, M.: UML Distilled: A Brief Guide to the Standard Object Modeling Lan-
guage. 3 edn. Addison-Wesley (2004)
2. Booch, G., Rumbaugh, J., Jacobson, I.: UML User Guide. Addison-Wesley (1999)
3. Group, O.M.: Unified Modeling Language Specification, version 2.0. Rational
Software Corporation (2003)
4. André, C., Rigault, J.P.: Variations on the semantics of graphical models for
reactive systems. In: System Management and Cybernetics, IEEE Press (2002)
5. Menezes, P.B., Sernadas, A.S., Costa, J.F.: Nonsequential automata semantics for
concurrent, object-based language. In: Electronic Notes in Theoretical Computer
Science - 2nd US-Brazil Joint Workshops on the Formal Foundations of Software
Systems. Volume 14., Elsevier (1998)
6. Barr, M., Wells, C.: Category Theory for Computing Science. Prentice Hall (1990)
7. Meseguer, J., Montanari, U.: Petri nets are monoids. Information and Computation
88 (1990) 105–155
8. Menezes, P.B., Costa, J.F., Sernadas, A.S.: Refinement mapping for general (dis-
crete event) system theory. In: Lecture Notes in Computer Science - 5th Interna-
tional Conference on Computer Aided Systems Theory and Technology. Volume
1030., Springer-Verlag (1996) 103–116
9. Reisig, W.: Petri Nets: an introduction. Volume 4 of Eatcs Monographs on Theo-
retical Computer Science. Springer-Verlag (1985)
10. Gehrke, T., Goltz, U., Wehrheim, H.: The dynamic models of UML: Towards a
semantics and its application in the development process. Technical Report 11/98,
Institut fur Informatik, Universitat Hildesheim (1998)
11. Eshuis, R., Wieringa, R.: Comparing petri net and activity diagram variants for
workflow modelling - a quest for reactive petri nets. In: Lecture Notes in Computer
Science - Petri Net Technology for Communication Based Systems. Volume 2472.,
Springer-Verlag (2003) 321–351
12. Shen, W., Compton, K., Huggins, J.: A validation method for uml model based
on abstract state machines. In: Proceedings of EUROCAST. (2001) 220 – 223
13. Knapp, A., Merz, S.: Model checking and code generation for uml state machines
and collaborations. In: Proceedings of 5th Workshop on Tools for System Design
and Verification. (2002) 59 – 64
14. Machado, J.P., Menezes, P.B.: Modeling transactions in uml activity diagrams via
nonsequential automata. In: Actas de la XXX Conferencia Latinoamericana de
Informatica, CLEI (2004) 543 – 553
15. Borger, E., Cavarra, A., Riccobene, E.: An ASM semantics for UML activity
diagrams. In: Lecture Notes in Computer Science - 8th International Conference on
Algebraic Methodology and Software Technology. Volume 1816., Springer-Verlag
(2000) 293–308
16. Rodrigues, R.W.S.: Formalising UML activity diagrams using finite state processes.
In: Workshop on Dynamic Behaviour in UML Models, 3rd International Conference
on the Unified Modeling Language. (2000)
17. Eshuis, R.: Semantics and Verification of UML Activity Diagrams for Workflow
Modelling. PhD thesis, University of Twente (2002)
Acknowledgments
This work was partially supported by CTXML/Microsoft/PUCRS, CNPq (Pro-
jects HoVer-CAM, GRAPHIT, E-Automaton) and FINEP/CNPq (Project Hy-
per-Seed) in Brazil.
One formal systemic approach in software design. Case
study: Extreme programming
Cestmir Halbich1
Abstract. The position paper comes out from Donald Schoen’s ideas about re-
flective practitioners and describes author’s experience in area of computer sci-
ence. The software gap is mentioned and methods for its overbridging too. At
the case study is illustrated Schoen’s approach to designer’s activity and benefit
for improving effectiveness and efficiency of the software design process by re-
flective practitioner’s approach. In the conclusion the advantages and disadvan-
tages are described.
1 Introduction
At the begining of 3rd millennium methods which help develop software products
become more important. They adjust it rapidly and flexibly to changing common
conditions. Often we call these methods agile methodologies. The idea of these is
simple. Only one way exists to check correctness of the developed system - the fastest
development as is possible and then to present results to the customer and adjust it
according by a feedback. Basic principles of agile methods are next:
- incremental and iterative development with very short iterations
- special emphasis on personal communication in frame of development team,
- permanent communication with user
- rigorose automated tests.
From historical point of view we know two approaches to system development. Hard
systems approach is based on systems engineering and systems analysis.The people
are treated as passive observers of the system development process. However, this
approach is not suitable in organizational environment that involves political, social,
or human activities. Development of such systems require an active involvement of of
every stakeholder. The approach that encompasses all the stakeholders of the system
is soft system approach. SSM is a methodology that adopts such ”soft system ap-
proach”.
The first is specific, with clear boundaries and parameters. Possible solutions are
implicit in the problem. The problem could be passed to a designer who could pro-
pose and implement a solution. The second is much less clear. It describes a problem
with weak boundaries and an unclear cause. In fact, it is not describing a problem, so
much as two symptoms. Until the symptoms are explored further, it is not possible to
say what the problem is, let alone propose a solution. Traditional systems analysis is
based on hard problems. These problems have a clear boundary and involve tangible
factors such as staff, documents, procedures, equipment, and structure. Its purpose is
to propose and design a solution for a predetermined problem.
Many business situations are less clear-cut, making a hard approach less appropriate.
Peter Checkland [1] developed Soft Systems Methodology (SSM), recognising that
many problems in organisations are affected by less tangible factors. These include
culture, informal interactions, and attitudes – which he terms the ”Human Activity
System”. SSM’s purpose is to investigate, understand and identify a problem. Its
investigation may reveal a number of problems to address rather than presuppose a
single root cause of difficulties.
SSM is based on systems thinking. It views the problem domain in a holistic rather
than reductionist way, recognising that the component parts are interconnected, so
that a change to one part will affect the other parts. Not only this, but the problem
domain itself is a subsystem of several larger systems – changes in one will affect our
domain as well.
Although Checkland coined the term Soft Systems Methodology, it is not strictly a
methodology. It does not prescribe a rigid series of steps that must be followed scru-
pulously. Rather, it suggests a framework for enquiry, with a number of recom-
mended techniques. The end result is an understanding of the problem domain so that
a hard study can then be applied to specify a solution.
version
of soft-
Action ware
to improve
comparison
Fig. 1. Scheme of the iterations and feedback loops in the software development by Extreme
programming.
3 Conclusion
References
1. Checkland, P.: Systems Thinking, Systems Practice, (1981), John Wiley & Sons, London
(1981)
2. Schoen, D. "Designing as reflective conversation with the materials of a design situation,"
Knowledge-Based Systems Journal - Special Issue on AI in Design, Vol. 5, No. 1, (1992),
3-14
3. NATO SOFTWARE ENGINEERING CONFERENCE 1968, Garmisch, (1968)
4. Beck, K.: Extreme Programming Explained: Embrace Change, 224 pages ; Addison-Wesley
Pub Co; 1st edition, (1999)
On the Application of Biological Vision Concepts and
Mechanisms in very Demanding Industrial Visual
Processing
Extended Abstract
C
A Acquisition Low Level Channel 1
M
Illumination 1
System
Text Presence Test
Low Level Channel n
ENE 05 08:25
L4185/01
2
Medium Level Channel n Text Irregularities Test
E$E0? 08:25 4
L4#85/0? Syntax Test
Code Validation
Semantic Test
Decision
Fig. 1. Multi-level and multi-channel architecture for validation
References
1 Alemán-Flores, M., Leibovic, K.N., Moreno Díaz jr, R.: A computacional Model for Visual
Size, Location and Movement, Springer Lectura Notes in Computer Science, Vol 1333.
Springer-Verlag. Berlin Heidelberg New York (1997) 406-419
2 Quesada-Arencibia, A., Moreno-Díaz jr, R., Alemán-Flores, M., Leibovic, K.N: Two
Parallel Channel CAST Vision System for Motion Analysis. Springer Lecture Notes in
Computer Science, Vol. 2178. Springer-Verlag. Heidelberg New York (2001) 316-327
3 Quesada-Arencibia, A.: Un Sistema Bioinspirado de Análisis y Seguimiento Visual de
Movimiento. Doctoral Dissertation. PhD Thesys. Universidad de Las Palmas de Gran
Canaria (2001)
4 J.C. Rodríguez Rodríguez, A.Quesada-Arencibia, R.Moreno-Díaz jr, and K.N. Leibovic: On
Parallel Channel Modelling of Retinal Processes Vol 2809 Springer-Verlag. Berlin
Heidelberg New York (2003) 471-481
5 Leibovic, K.N., Science of Vision, Springer Verlag, New York, 1990.
Theory-Building with System Dynamics: Principles and
Practices
It is crucial for the long-term success of all kinds of enterprises, including orchestras,
to discover and take opportunities as early as possible as well as to detect and avoid
potential threats before they become uncontrollable. Therefore, the anticipative
capturing of future realities by being able to interpret so-called weak signals regarding
external developments is important and might be critical for the long-term survival of
an organization (Ansoff 1975). Regarding the orchestra sector such a weak signal lies
in the fact that the German and Swiss classical music audience has a disproportionate
number of old people when compared to the population as a whole (significance level:
p = 0.001). The implications for the future size of the classical music audience and the
related demand for classical music life performances remained unclear, since there
was no theory about the formation of an individual’s affinity with classical music. In
particular, there was disagreement as to whether the fundamental musical taste is
predominantly subject to an aging/life cycle effect or a cohort effect due to the lack of
a theory. If people mature into listening and attendance practice as they become older
or by entering a new stage in the course of life (aging/life cycle effect), the classical
music audience will increase in the next decade. But if the younger adults with
decreased participation will not change their musical preferences over time (cohort
effect), the classical music audience will gradually die out. In order to fill this
research gap, one of the authors developed a theory about how the individuals’
musical preferences for classical music come into existence, by drawing on the SD
methodology (Hamann 2005).
4 On Model Validation
5 Conclusions
Extended Abstract
Statistical information plays a central role in many business and economic decisions.
The term information means here that we have to consider the data itself as well as
descriptions of the data, so called metadata, which are necessary for obtaining the
information. Hence, it is not surprising that metadata play a key role in statistical
information systems for a long time. The earliest reference is Sundgren [11], who
introduced the concept of metadata in statistics for a so called infological approach
towards statistical information. This approach has been developed further in many
ways by a number of researchers as well as statistical agencies and has lead to a
number of metadata standards. Due to the fact that a lot of statistical information is
contained in highly aggregated data, represented nowadays usually in data
warehouses, one line of development focussed on metadata standards for such type of
data (for example the SDDS-standard [9]). A second stream of development based on
ideas from documentation for data archives considered mainly the metadata
requirements for scientific researchers in social sciences and economics. These efforts
have resulted in different metadata standards, probably the best known example is the
DDI standard [1], which is a substantive expansion of the Dublin Core metadata [2].
A well known software tool based on these ideas is NESSTAR [7]. A further
development concentrated on proper metadata representation for value domains of the
attributes of interest and lead to the so called “Neuchatel Terminology” [8] for
classifications.
Due to the different starting points of these approaches it is rather cumbersome to
integrate data in cases where the definition of the data schemes is based on such
different documentation schemes. Following the developments of intelligent
information processing in recent years (see for example Sowa [10]) also in the field of
statistical information processing there were a number of efforts to develop the idea of
metadata further into the direction of ontology. In fact, statistical metadata practice
includes to a far extent information needed for ontology – engineering. Probably the
most important contribution in this direction was made by the METANET project (cf.
[4, 6]), a thematic network within the fifth EU-research framework. The approach
tries to fulfil the requirements of the ontology definition of Gruber [5] (“ontology” is
a specification of a conceptualization) by formalizing the statistical data paradigm,
taking into account the representational, operational and functional semantics of
statistical information.
Starting points are basic statistical objects like dataset, statistical population, or
statistical variable, which constitute the categories for the ontology. For all these
categories a unified description framework was developed, which is called the view
facet of statistical categories. The following four views were distinguished:
• The conceptual category view represents the subject-matter definition of any
category instance and builds the bridge to reality. Validity of the definition of the
subject matter concept gets usually restricted by temporal and geospatial
constraints.
• The statistical (methodological) category view describes the statistical properties
of the category instance by using a number of parameters, which have to be
defined in such way that specific properties of the different categories are taken
into account.
• The data management category view is geared towards machine supported
manipulation storage and retrieval of the category instance data.
• The administrative category view addresses management and book-keeping of the
structures.
Based on these view facets a representation scheme can be defined, which seems to
be sufficient for operational purposes. A first sketch of such a model was presented in
Denk et al. [3]. In this paper we present first and fundamental ideas for using this
framework for mapping different metadata standards.
References
Arminda Moreno-Dı́az1 , Miguel Virto1 , Jacinto Martı́n2 , and David Rı́os Insua3
1
School of Computer Science. Madrid Technical University.
{amoreno, mvirto}@fi.upm.es
2
Dept. of Mathematics. University of Extremadura.
jrmartin@unex.es
3
Statistics and Decision Sciencies Group. Rey Juan Carlos University.
drios@escet.urjc.es
References
1. Demirer, R., Charnes, J.M., Kellogg, D.: Influence Diagrams For Real Options Va-
luation. Working Paper No. 03-0805. SUIE School of Business (2003)
2. French, S., Rı́os Insua, D.: Statistical Decision Theory. Arnold (2000)
3. Howard, R.A., Matheson, J.E.: Influence Diagrams. In: R.A. Howard and J.E. Mat-
heson (eds.): Readings on the Principles and Applications of Decision Analysis. Vol.
2. Strategic Decision Group. Menlo Park (1984) 719–762
4. Lander, D., P.P.Shenoy.: Modeling and Valuing Real Options using Influence Dia-
grams. Working Paper No. 283. School of Business. University of Kansas. Lawrence,
KS (1999)
5. Nease, R.F., Dopuglas, K.O.: Use of Influence Diagrams to Structure Medical De-
cisions. Medical Decision Making 17 (1997) 263–276
6. Pearl, J.: Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann (1988)
7. Stafira, S. Jr., G. S. Parnell, J. T. Moore.: A methodology for evaluating military
systems in a counterproliferation role. Management Science 43 (1997) 1420-1430
8. Tanner, M.A., Wong, W.: The Calculation of Posterior Distributions by Data Aug-
mentation (with discussion). Journal of the American Statistical Association 82
(1987) 528550
9. Virto, M.A., Martı́n, J., Rı́os Insua, D., Moreno-Dı́az, A.: Approximate Solutions
of Complex Influence Diagrams through MCMC Methods. In: Proceedings of First
European Workshop on Probabilistic Graphical Models. Gámez, Salmerón (eds.)
(2002) 169–175
10. Virto, M.A., Martı́n, J., Rı́os Insua, D., Moreno-Dı́az, A.: A Method for Sequential
Optimization in Bayesian Analysis. Bayesian Statistics, 7. J.M. Bernardo, M.J. Ba-
yarri, J.O. Berger, A.P. Dawid, D. Heckerman, A.F.M Smith, M. West (eds.) Oxford
University Press (2003) 701–710
Decision support system using data warehousing
and on-line analytical processing technologies
Extended abstract
References
Extended Abstract
Starting the planning agent offline we calculate our first project plan, con-
taining a complete workflow of all activities and schedules of all involved persons
(Mauerkirchner, 2001). Subsequently both systems (real system and simulation
system) are ready to be started simultaneously based on our plan. In general, all
feedbacks stem from the real project system and enforce normally a recalcula-
tion of the actual project plan and, based on it a synchronisation of real system
and simulation system Feedbacks can be categorized in general feedbacks (e.g.
assignments or temporary absence of human resources) and personal feedbacks.
The latter ones come from persons using their human agent’s assessment inter-
face. Messages of isolated persons (groups with only one member) are sent to
the controlling agent at least (decision of project manager is necessary), while
feedbacks of other groups are received by the interference agent . It is possible to
create useful updates to the set of characteristic properties used by the central
planning agent. The bases for these updates are generated by extracting rele-
vant information from suggestions for improvement made by the human users
of the real system. These suggestions are transferred by the personal agent´s
assessment interface to the interference agent and stored as fuzzified statements
containing comments on problem relevant user skills and critical member inter-
actions in each user group. The collection of these statements is used by the
interference agent to generate an update of the above mentioned system prop-
erties in following distinct steps :
– create a population of sets that encode one statement for each characteristic
property, respectively.
– determine the fitness of each set by repeating offline the latest simulation of
the planning agent.
– choose sets that are able to improve the synchronisation of the actual project
flow in the real system and the simulated one.
– generate new populations of sets by performing crossover and mutations.
– repeat the last three steps until some improved steady state is reached.
The updated property values are checked by the controlling agent for plausi-
bility and then transferred to the planning agent for future simulations of project
flow. An alternative to this procedure might be the utilisation of Baysian Belief
Networks (Druzdel, 1996)(Pearl, 1988).
References
1. M.J.Druzdel, (1996) : Qualitative Verbal Explanations in Baysian Belief Networks.
AISB Quarterly1.96
2. J.Pearl, (1998): Probabilistic Reasoning in Intelligent Systems: Networks of Plausi-
ble Inference. Los Altos, CA: Morgan Kaufmann
3. M.Mauerkirchner, (2001): A General Planning Method for Allocation of Human Re-
source Groups, in Lecture Notes in Computer Science- Proceedings EURO-CAST’01
Las Palmas de Gran Canaria, Spain
On Recursive Functions and Well{founded
Extended Abstract
?
1 Introduction
Let a binary relation on a set A. The set Acc(A; ) of accessible elements of
is the set of elements a 2 A such that there is no innite descending sequence
n+1
:::a an :::a 2 a1 a:
The relation A A is well{founded if
A = Acc(A; ):
Formally we dene Acc(A; ) as a recursive predicate on A : an element x is
accessible if every y x is accessible.
Inductive Acc [A:Set;R:A->A->Prop]:A->Prop:=
Acc_intro:(x:A)((y:A)(R y x)->(Acc A R y))->(Acc A R x).
Definition well_founded := [A:Set;R:A->A->Prop](a:A)(Acc A R a)
In the paper detailed proofs of the justication of this representation are
given via informative versions of well{foundedness. Also, some connections with
intuitionistic and classical logic aspects of the models are analyzed.
? Government of Galicia, Spain, Project PGIDT02TIC00101CT and MCyT,
Spain, Project TIC 2002-02859, and Xunta de Galicia, Spain, Project
PGIDIT03PXIC10502PN
2 Results and Methods
Beginning from signicative examples we generalize to obtain some methods that
seem to be the base of a practical methodology to dene programs by general
recursion. This new methods are illustrated with some non trivial examples.
References
1. A. Abel, foetus Termination Checker for Simple Functional Programs. Techni-
cal Report. Theoretical Computer Science, Institute of Computer Science, Ludwigs{
Maximilians{ University (1999)
2. A. Abel, Termination Checking with Types. EDP Sciences (1999)
3. A. Balaa, Fonctions recursives generales dans le calcul des constructions. PhD.
Theese. Universite de Nice{Sophia Antipolis (2002)
4. A. Bove, Simple General Recursion in Type Theory. Technical Report. Chalmers
University of Technology, Goteborg (2000)
5. G. Doweck, A. Felty, H. Herbelin, G. Huet, C. Murty, C. Parent, C. Paulin{Mohring,
B. WernerThe Coq porrf assistant user's guide. Technical Report 134, INRIA. (1993)
6. B. Nordstrom, Terminating General Recursion. BIT, Vol. 28 (1988)
!
" #$%
&'%
(
!
)
(
*
+ +,
++" *
-
.
/012
!
&& -
3
(
"
+
4
2
5
,
"
(
!
6&7 "
,
*
*
,
,
*
(
(
* 4
8
(
* "
9
,
, ,
89
(
)
(
*
)*
!
6&7
*
)* !
4 (
0
)* )*"
++" :
(
;)*"
;
)*"
:
;
;
)*"
(
;4 )*" (
;
(
.
)*" .
+
<
. $
"
<<
)*
++"
2
,
)*"
.
)*
++"
)*"
!
,
+"
+ ++"
(
/
,
.
/
&
&
/
"
/
,
"
&
.
/ +
/
,
,
&
= &
/67
/6 7>
/ ,
?
5 (
+ +,
@
,
.
,
++=
,
&
+ +,
&
@
,
.
/ 0$ !
,
6& &1 $& $0 017
& '
( )
* +,
-
,
%
"
!. !/ &
0
%
1
02
# 3
# 45 %
6 3# $ 7
( $$89
9
# :9
#;
! << """&"=&">? "'$ 44+4 "$ +&5 @4"(1 44+4A
93# 7BC D E 9%, 3# D E 9%, 3#
, D E
89 D E 9 9
E 9FF;
G 93# 7B9C D E << 4&1 44+4 $&'$ @"1 @4"(1
H
89I D E 9 9 E 9IFF; << >' & )'4J44$$>'$ > & 44+4
/
<< @4 142 14 @4"(1 > &$$"( > 1"$ 44+4
89B9 C 9B9 C ::
9
B9 C D 9
B9 C ;
!
93# 7B9C D 93# 7B9IC F E
G
H << @4 244) '&2 > 14 44+4 @"1 1&$ +&5"++ @4"(1
/ 89
B9 C D 9
;
9%, 3# D 93# 7B9CE
9%, 3#
, D 9E
!
<< +K4' > 44+4$ " $4L44 D +&5 @4"(1
9
# D 9%, 3#E
G 9 D MME
H 9
E << &''&N$ $&' & @4"(1$ $&' & O
/ 9B9
C D 9B9
CE
! 9 D 9%, 3#
, 0 E
! 3# 8 9
; << @1"4 > & @4"(1$ >P4'4J
!
!! 893# 7B9C DD 9%, 3#;
!
! 9%, 3#E
! 9B9%, 3#C D 9
B9CE
!G
!H 9 E
!/
9E
<< "> 4J$ 14'4
$$
#%
Information Retrieval and Large Text
Structured Corpora ?
Extended Abstract
Conventional Information Retrieval Systems usually deal with plain text docu-
ments or with data files with a very elementary structure. Therefore, these kinds
of system are able to solve queries in a very efficient way, but only taking into
account whole documents, without distinguishing between different sections in
those documents. This behaviour makes it impossible to support queries that
should throw up a subset of documents which have some common characteristic
or perform searches based on document sections.
In contrast with this, text corpora commonly employed nowadays are usually
composed of a set of text files which in some way show a complex structure. So,
building a classical Information Retrieval System to work with this kind of data
will not benefit from this structure and results will not be improved.
In this article we test different technologies which can be used to build In-
formation Retrieval Systems to be applied on large structured corpora, namely,
technologies able to use the document structure to offer possibilities of filtering
the search. In addition, and as a result of our work, we can propose the technol-
ogy with the best flexibility-performance balance for use in building this kind of
system.
There are several examples of Information Retrieval Systems using struc-
tured corpora [1] [2], and many others about searching methods on large text
corpora [3], but there is no analysis or comparative study on technologies and
architectures that can be used to build them. An overview of different generic
technologies that could be applied to XML Information Retrieval Systems is
shown in XML Data Management, Native XML and XML-Enabled Database
Systems [4], but it is not about corpora.
In our analysis we suppose that the considered corpora are organized as a set
of different XML (Extensible Markup Language) [5] files. This assumption does
?
Partially supported by Ministerio de Educación y Ciencia (MEC) and FEDER
(TIN2004-07246-C02-01 and TIN2004-07246-C02-02), by MEC (HF2002-81),
and by Xunta de Galicia (PGIDIT02PXIB30501PR, PGIDIT02SIN01E and
PGIDIT03SIN30501PR)
not imply an important restriction since every text corpora can be transformed to
this format either automatically or manually, depending on the original structure
of the text.
Tamino 1 (a native XML indexer) [6], and Oracle2 [7] (a relational database
system which in its last versions integrates XML capabilities) are the main alter-
natives taken into account in our work. To perform our tests, we have taken into
account the following two main criteria: query flexibility and query performance.
Query performance evaluation is based on processing time measures using a
set of representative queries that covers a wide range of questions performed by
typical text corpora users. To evaluate query flexibility we have considered the
following aspects of the technologies under review:
The results obtained show that although XML native managers are a serious
candidate for the near future, they present certain deficiencies, which limit query
flexibility if we want to achieve a good performance. Only high performance
relational database systems like Oracle can satisfy some kind of requirements,
relevant to this kind of system.
1
Tamino is a Software AG product.
2
Oracle is a registered trademark of Oracle Corporation and/or its affiliates.
References
1. Marı́a Sol López Martı́nez CORGA (Corpus de Referencia del Gallego Actual) In
Proc. of Hizkuntza-corpusak. Oraina eta feroa, pages 500–504, Borovets, Bulgaria,
Sept. 2003.
2. Mark Davies Un corpus anotado de 100.000.000 palabras del español histórico y
moderno. In Proc. of Sociedad Española para el Procesamiento del Lenguaje Natural
pages 21–27, Valladolid, Spain, 2002.
3. Mark Davies Relational n-gram databases as a basis for unlimited annotation on
large corpora. In Proceedings from the Workshop on Shallow Processing of Large
Corpora, Lancaster, England, pages 23–33, Lancaster, England, March 2003.
4. Akmal B. Chaudhri, Awais Rashid and Roberto Zicari XML Data Management,
Native XML and XML-Enabled Database Systems, Addison-Wesley, March 2003.
5. XML In http://www.w3c.org, 27/10/2004
6. Tamino In http://www.softwareag.com, 27/10/2004
7. Oracle In http://www.oracle.com, 27/10/2004
Meteorological Image Descriptors
1 Introduction
The main aim of this paper is to obtain a visual characterisation of time evolution im-
ages, in particular, synoptic maps taken from Meteorology. As far as we know, no
similar work has been found to date in the computer vision literature for classification
or prediction of this kind of image, besides those that treat the problem with compli-
cated numerical models in Meteorology or statistical mathematical methods.
The preprocessing phase of the synoptic maps that we are working with has al-
ready been explained in [1]. Common preliminary tasks to single out the image in-
formation of interest, such as useless background removal, information separation and
memory usage have been discussed there.
The next step is studied in this paper and consists in extracting numerical descriptors
of the images in order to obtain a visual characterisation of them [2].
Since we are dealing with meteorological images, we focus on features that are re-
lated to atmospherical properties. In this sense, two types of numerical descriptors are
extracted: the low level and high level ones. The latter will be subsequently used for
prediction tasks, while the former will be used for classification tasks.
The low level numerical descriptors are the density of the isobar lines, their average
orientation in local areas, and the location of high (H) and low (L) pressure points.
The density of the lines is actually a downscaled image, following the approach of
pyramid representations [3]. For this purpose, we consider the isobar image to be di-
vided into adequate small windows. On the one hand, the size of the windows must be
small enough to identify the corresponding isobar image. However, on the other hand,
the window must be as big as possible in order to minimize the memory used per iso-
bar image, which will be important for the subsequent analysis, given the high num-
ber of images to be processed and stored.
The detection of H and L letters is a typical OCR task, but since no other letters are
involved, we use an ‘ad-hoc’ algorithm, instead of generic OCR techniques [chapter 2
of 4]. Simple methods do not work, because H and L letters present in our images are
not equal. For this reason, an analysis of the discriminatory power of several features
has been carried out.
To verify if the aforementioned extracted variables are good low level numerical
descriptors, we develop the inverse procedure, using them as the input for rebuilding
the synoptic map, and compare it with the original one. Satisfactory descriptions are
obtained for adequate window width.
Regarding the high level numerical descriptors, two main features are represented
per image: the big regions of locally profound high (H) and low (L) pressure and
those of nearly constantly oriented circulation.
Acknowledgments
The authors are deeply grateful to the Spanish Interministerial Board for Science and
Technology (Comisión Interministerial de Ciencia y Tecnología CICYT) which sup-
ports Project TIC2002-01306, as a part of which this abstract has been produced.
They would also like to thank the Meteorological Centre in EUVE (European Virtual
Engineering) Technology Centre (Vitoria, Spain) for their collaboration.
References
1. Crespo, J. L., Bernardos, P., Zorrilla, M. E., Mora, E.: Preprocessing Phase in the PIETSI
Project (Prediction of Time Evolution Images Using Intelligent Systems). 9th International
Workshop on Computer Aided Systems Theory: Eurocast 2003. Eds. Roberto Moreno-Díaz
and Franz R. Pichler. Lecture Notes in Computer Science 2809, 651-660. Springer-Verlag,
Berlin Heidelberg New York (2003)
2. Sonka, M., Hlavac, V., Boyle, R.: Image processing, analysis and machine vision. Interna-
tional Thomson Computer, London, Madrid (1996)
3. Shalkoff, R. J.: Digital Image Processing and computer vision. Wiley, New York (1989)
4. Cole, R. A. (ed. in chief): Survey of the State of the Art in Human Language Technology.
Cambridge University Press (1996)
Matching with a Hierarchical Ontology
—Extended Abstract—
1 Introduction
2 Hierarchical Ontologies
For classifying and scoring product descriptions, as well as queries, we make use
of declarative hypernym/hyponym (is-a) and meronym/holonym (part-of ) hier-
archies, such as the sense hierarchies provided by WordNet [3]. We assume these
ontologies are organized as directed acyclic graphs. Each node—representing
a class of objects (commodities and commodity categories, in our case) or an
attribute—may be related by is-a and part-of edges to other nodes. For exam-
ple: a rucksack is-a kind of backpack, which is-a kind of bag (in the sense of
container), jet black is-a shade of the attribute black, and a clasp is a part-of a
handbag, as is a handle.
Fig. 1. Sample search dialogue.
3 Classification
Making use of a hyponym taxonomy has long been a tool in hierarchical classi-
fication. See, for example, [5, 2, 6].
There are several novel ways in which an ontology can help:
For example, consider a product labelled “Totes Hooded Zip Front Jacket”
(“Totes” is a brand name) and described as follows: “This spring parka features
a zip front, two front pockets, a hood with a drawstring and an inside contrasting
pattern. The jacket is ideal for those warm, rainy days and is both functional
and fashionable.” Linguistic analysis suggests it may be a jacket (appears in the
name as well as in the catalog description) or a parka (appears in the beginning
of the descriptive paragraph). Since parka is a hyponym of jacket, the former,
more specific, category is preferred.
Also, the mention of pocket, a meronym of garment, which is a (distant) hy-
pernym of jacket, supports both interpretations. Since pocket is not a commodity,
it is not itself a contender. Furthermore, the words “parka” and “pocket” both
reinforce the garment sense of “jacket.”
Ambiguous queries may be subjected to similar analysis.
4 Retrieval
1. Retrieve all hyponyms (e.g. rucksacks and etuis) of query terms (“bag”),
but not their hypernyms.
2. If the query (“orange leather briefcase”) is for a subcategory in which there
are too few elements, one may wish to suggest siblings (orange leather back-
packs; red leather briefcases) or cousins from the hypernym hierarchy.
3. In some contexts, a query for a category might also propose its meronym
classes (lens for “cameras”) as also of relevance.
5 Results
We thank Michael Flor for his large share in this effort and Yael Villa for her
comments and encouragement.
References
1. Hao Chen and Susan Dumais. Bringing order to the Web: Automatically catego-
rizing search results. In Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems, pages 145–152, The Hague, The Netherlands, April 2000.
ACM Press.
2. Susan T. Dumais and Hao Chen. Hierarchical classification of Web content. In
Nicholas J. Belkin, Peter Ingwersen, and Mun-Kew Leong, editors, Proceedings of
SIGIR-00, 23rd ACM International Conference on Research and Development in
Information Retrieval, pages 256–263, Athens, GR, July 2000. ACM Press.
3. Christiane Fellbaum, editor. WordNet: An Electronic Lexical Database. MIT Press,
Cambridge, MA, 1998.
4. Tal Rubenczyk, Nachum Dershowitz, Yaacov Choueka, Michael Flor, Oren Hod,
and Assaf Roth. Search engine method and apparatus. Patent pending.
5. Sam Scott and Stan Matwin. Text classification using WordNet hypernyms. In Pro-
ceedings of the COLING/ACL Workshop on Usage of WordNet in Natural Language
Processing Systems, pages 45–52, Montréal, Canada, August 1998.
6. Aixin Sun, Ee-Peng Lim, and Wee-Keong Ng. Hierarchical text classification meth-
ods and their specification. In Alvin T. S. Chan, Stephen C. F. Chan, H. V. Leong,
Vincent, and T. Y. Ng, editors, Cooperative Internet Computing, pages 236–256.
Kluwer Academic, Boston, March 2003.
Business Intelligence Process Through Information
Extraction with Lixto
Extended Abstract
Key Intelligence
Topic
Planning and
directions
Decision Makers
Data
Dissemination Collection
Analysis and
Other Reporting
Users
Finally, structure data corresponding to the output type(s) must be extracted from
the answer HTML document wich contains irrelevant data as well.
A simple extractor processes an HTML document and constructs objects
corresponding to some output type specified in the corresponding row of the Wrapper
Capability Table. The contents of the HTML document are represented as an ordered
tree , so that declarative queries may be written to describe the data that is to be
extracted from the document. These queries use qualified path expressions to identify
nodes based on their contents; to traverse the ordered tree; and to extract data from the
elements (nodes). Additional operators of this languages provide the control to
manipulates data and to construct the output objects.
The contents of an HTML document and its structure (placement of the HTML
elements) are represented as an ordered tree, with the document as root. The nodes of
the tree correspond to the HTML elements of their components, e.g., table, list, img,
table row, list element, etc. The arcs of the tree represent the placement of the HTML
elements of their components within the document. The preorder enumeration of the
ordered tree corresponds to the top-down placement of HTML elements and their
components in the document. In a similar way to WebOQL [5], we represent an
HTML document as an ordered tree. However in contrast to WebOQL hypertrees, in
our representation each node is associated with a type wich represents an HTML
element or its components. The arc relate an HTML element with the list of HTML
elements embedded in it or the parameters that modify it. Our representation of an
HTML document allows us to simplify the query language of our extractors.
Consider of an HTML document whose ordered tree representation . It is
comprised of an HTML list followed by two HTML tables, each of wich is a node.
The components of these elements are also a nodes, e.g., table rows, (tr), and table
data (td). One td element comprises of an img element, wich is also a node.
In our data model, each HTML document (HTML element or parameter) is
described by a node in the ordered tree with the following attributes:
• Name: This represent the type of HTML element or its component, e.g., table, tr,
td, img, src, etc.
• Ocurrence: This is an integer representing the placement of this object in the
HTML document. The placement is recorder w.r.t. the HTML element that is the
parent node. The intergert value indicates the placement of this node w.r.t. the
sibiling node of the same type. In the figure 3. 4 the second table object has an
occurrence of 2, indicating it is the second table, even through it is the third child
of the parent (root).
• Data: This represent the optional data associated with an HTML element. The data
is particular to the type of element, for e.g., for the HTML element img, one
component is the scr parameter whose data is usually a URL.
• Child: This is a relationship that relates the object (HTML element or component)
represented by the node to the list of objects embedded in it or to the parameters
that modify it. It represents the arcs between a node and its children in the ordered
three. For example, the second ocurrence of the HTML element table in figure 3,
has two children of type tr. The second tr has one child of type, wich has one child
of the type img. The parameter src is represented as a child of the node img.
HTML document
root
List 1
Table2 Table1
Li 1 Li 2 tr 1 tr 2 tr 1 tr 2
td 1 td 2 table td 1
tr 1 tr 2 Img 1
src 1
td 1 td 1
With this kind of representation our data extraction model will take sense and we
can do the intelligence process to take the relevant information of each of the websites
in XML, with these information we will feed the databases and we will can to develop
a corporate repository related with the transactional sytems in the organization.
References
LFCIA Lab.
Department of Computer Science,
Universidade da Coruña
Campus de Elviña, s/n, 15071 A Coruña (Spain)
{sjorge,gulias,freire,juanjo}@dc.fi.udc.es,
http://www.lfcia.org
Extended Abstract
Proving the correctness of a program, even a simple one, is a complicated task and at
the same time it is one of the most important activities for a software engineer.
Functional programming [6,1,11] has often been suggested as a suitable tool for
writing programs that can be formally analyzed, and whose correctness can be as-
sured [5,8,7]. That assertion is due to referential transparency, a powerful mathematical
property of the functional paradigm that assures that equational reasoning, i.e., f x is al-
ways f x, makes sense. As a matter of fact, the mathematical way of proving theorems
can be successfully applied to computer science.
On the other hand, theorem provers help us in proving the correctness of a program.
These tools provide control and assistance in the processes of specification and prove;
thus they prevent bugs that could be introduced in an hand-made proof. The logical
framework Coq [3] is an implementation of the Calculus of Inductive Constructions
(CIC) [2,10], performed at INRIA. This system is a theorem prover goal-directed and
tactic-driven in the style of LCF [4].
The aim of this work is to contribute to the construction of a methodology to produce
certified software, i.e., we intend to find the way to strengthen two different notions:
First, there is a need of formal verification of the correctness of algorithms which goes
together with their construction; and second, programs are mathematical objects which
can be handled using logico-mathematical tools. These methodological techniques will
be exemplified by means of a program which calculates the Gröbner base of a set of
multivariate polynomials. We present the way in which one can formally verify several
aspects of the application implemented following the functional paradigm.
In this work, both the certification of the program and its efficiency will be taken
into consideration.
We present the proofs of properties in an informal (but exhaustive) style, with man-
ual proofs, where each step is justified by means of mathematical reasoning; and also,
the formalization of these proofs in the Coq proof assistant.
? This work has been partially supported by MCyT TIC 2002-02859 and Xunta de Galicia
PGIDIT03PXIC10502PN
First, we create the bases of the mechanization of an algebraic theory for the rings
of multivariate polynomials using dependent types.
Afterward we develop a theory about well-founded relations. A well-founded poly-
nomial ordering is formalized in the proof assistant Coq. Our aim here, is to establish
the accurate bases for proving the termination of certain reduction relations of polyno-
mials.
To conclude, an application in the field of symbolic computation is both developed
and verified: the calculation of the Gröbner basis for a set of multivariate polynomials.
We face a challenge: Reasoning about computer algebra programs requires the formal-
ization of an important amount of non-trivial mathematics.
We intend not only the formal certification of certain aspects of the program written
in the functional language Caml [9] but also the performance of efficient computing.
References
1. R. Bird and P. Wadler. Introduction to Functional Programming. Prentice Hall, 1987.
2. T. Coquand and G. Huet. The Calculus of Constructions. Information and Computation,
1988.
3. G. Dowek, A. Felty, H. Herbelin, G. Huet, C. Paulin-Mohring, and B. Werner. The COQ
Proof assistant user’s guide. Technical report, INRIA, 1991.
4. M. J. Gordon, A. J. Milner, and C. P. Wadsworth. Edinburgh LCF: A Mechanised Logic of
Computation. Springer-Verlag, 1979.
5. V. M. Gulías, J. S. Jorge, , J. J. Quintela, and J. L. Freire. From Sequential to Distributed
Formal Proofs. In F. Pichler and R. Moreno-Díaz, editors, Computer Aided Systems The-
ory (EUROCAST’97), volume 1333 of Lecture Notes in Computer Science. Springer-Verlag,
1997.
6. P. Hudak. Conception, Evolution, and Application of Functional Programming Languages.
ACM Computing Surveys, 1989.
7. J. S. Jorge. Estudio de la verificación de propiedades de programas funcionales: de las prue-
bas manuales al uso de asistentes de pruebas. PhD thesis, Departamento de Computación,
Facultade de Informáticas, Universidade Da Coruña, oct 2004.
8. J. S. Jorge and J. L. Freire. Prove and then Execute. In 7o Encuentro de Álgebra Computa-
cional y Aplicaciones (EACA-01), 2001.
9. X. Leroy. The Objective Caml System: Documentation and User’s Manual, 2001. V3.04.
10. C. Paulin-Mohring. Inductive definitions in the system Coq: Rules and properties. In
M. Bezem and J. F. Groote, editors, Typed Lambda Calculi and Applications (TLCA’93),
volume 664 of Lecture Notes in Computer Science. Springer-Verlag, 1993.
11. L. C. Paulson. ML for the Working Programmer. Cambridge University Press, 2nd edition,
1996.
CheapTB: A Low Cost of Operation Distributed
File System
There are several applications that have been developed in recent years which
need great amounts of storage space, often bigger than what a single worksta-
tion can provide. Examples of this applications are the storage of digital video,
scientific experiments data or weather forecasting.
The traditional way of solving this problem is the use of some costly storage
service, like a SAN. Business which sell those systems often include maintenance
so this is an adequate solution for companies and research centers without budget
constraints. However, this big costs make it difficult for small business or research
labs to use this sort of solutions as they cannot afford them. The objetive of
CheapTB is to offer a distributed filesystem whose focus is low cost of adquisition
and operation rather than high performance.
As the storage capacity of IDE and SCSI disk grows it is increasingly possible
to put several disks in one machine. A cluster of these machines can be used as a
storage service. The cost of a storage cluster like the one described is much lower
than that of a SAN solution, although performance will be lower. Still, there
are many applications which need high storage space that don’t have critical
requirements in throughput or latency.
Some software is needed to use all of the space offered by the disks as if it
was a single disk. This software must manage the storage of files in the cluster
distributing them between the different disks the machines have, in several steps:
As this software controls the way data is stored in the cluster, it can control
power usage by the system powering off disks and machines which are not in use
at a certain time. This makes the system cheaper to operate.
Although it is possible to shut down machines in the cluster which are not in
use, at least a control node must always remain on to answer to external requests.
This node will be the same that translates external requests into accesses to the
different nodes, as it must know where all the data is in order to know which
machines must be powered on to answer an external request.
The system has an structure in three layers. The first layer is the interface
with the user. This layer offers access to CheapTB through several common
protocols, like NFS or SMB, using protocol adapters. CheapTB is designed so
that it is easy to add new protocols as needed.
These protocol adapters work on top of a common API with the second layer,
which is the filesystem. The API is similar to POSIX[1], to make adaptation of
existing network filesystem servers easy.
The filesystem manages the storage of the data and keeps track of the direc-
tory structure. The data is stored in a virtual block device, which is the third
layer. This virtual block device hides the complexity of managing all the different
storage media and the distributed nature of the system from the upper layers.
The power management works by distributing the responsability between the
filesystem, which has all the information about files and directory structure, and
the virtual block device, which has all the information about devices and nodes.
References
1. IEEE, ”Institute of Electrical and Electronics Engineers”, Standard 1003.1, 2003
Edition http://www.unix.org/single unix specification/
2. Sun Microsystems, Inc., NFS: Network File System Protocol Specification
http://www.ietf.org/rfc/rfc1094.txt
3. B. Callaghan, B. Pawlowski, P. Staubach and Sun Microsystems, Inc., NFS Version
3 Protocol Specification, http://www.ietf.org/rfc/rfc1813.txt
4. Christopher R. Hertel, Implementing CIFS, http://www.ubiqx.org/cifs/
Spelling Correction on Technical Documents?
1 Introduction
2 Asymptotic behavior
1e+06 1e+07
Our proposal Our proposal
Savary’s proposal Savary’s proposal
Number of computed items
1e+06
100000
100000
10000
10000
1000
1000
100
100
10 10
0 2 4 6 8 10 12 14 16 16 14 12 10 8 6 4 2 0
Length of the current suffix Position of the first point of error
1
Our proposal Our proposal
Savary’s proposal 1,4 Savary’s proposal
Number of computed items
Finally, the regional (resp. the global) approach provided as correction the
word from which the error was randomly included in a 77% (resp. 81%) of
the cases. Although this could be interpreted as a justification to use global
methods, it is necessary to remember that we are now only taking into account
morphological information, which has an impact in the precision for a regional
approach, but not for a global one that always provide all the repair alternatives
without exclusion. So, the consideration of the precision concept represents, in
the exclusive morphological context considered, a clear disadvantage for our
proposal since it bases its efficiency in the limitation of the search space. We
attend that the integration of linguistic information from both, syntactic and
semantic viewpoints will reduce significantly this gap, less than 4%, around the
precision; or even will eliminate it.
References
1. Thorsten Brants. Some experiments with the CRATER corpus. Technical report,
Universität des Saarlandes, Saarbrücken, 1995.
2. J. Graña, F.M. Barcala, and M.A. Alonso. Compilation methods of minimal acyclic
automata for large dictionaries. Lecture Notes in Computer Science, 2494:135–148,
2002.
3. K. Kukich. Techniques for automatically correcting words in text. ACM Computing
Surveys, 24(4):377–439, December 1992.
4. A. Savary. Typographical nearest-neighbor search in a finite-state lexicon and its
application to spelling correction. Lecture Notes in Computer Science, 2494:251–
260, 2001.
Verification of Language Based Fault-Tolerance
Extended Abstract
3
http://homepages.cwi.nl/ mcrl/mutool.html
4
http://www.inrialpes.fr/vasy/cadp/
References
1. Armstrong, J., Virding, R., Wikström, C., Williams, M.: Concurrent Programming
in Erlang (Second Edition). Prentice-Hall International (UK) Ltd. (1996)
2. Wiger, U.: Four-fold increase in productivity and quality – industrial-strength func-
tional programming in telecom-class products. In: Proceedings of the 2001 Workshop
on Formal Design of Safety Critical Embedded Systems (FEmSYS 2001). (2001)
3. Torstendahl, S.: Open telecom platform. Ericsson Review 1 (1997)
4. Arts, T., Earle, C.B., Penas, J.J.: Translating Erlang to mucrl. In: Proceedings
of the International Conference on Application of Concurrency to System Design
(ACSD2004), IEEE Computer Society Press (2004)
5. Groote, J.: The syntax and semantics of timed µCRL. Technical Report SEN-R9709,
CWI, Amsterdam (1997)
6. Arts, T., Earle, C.B., Derrick, J.: Development of a verified Erlang program for
resource locking. International Journal on Software Tools for Technology Transfer
(STTT) 5 (2004) 205–220
Applying Stacking and Corpus Transformation
to a Chunking Task
1 Extended abstract
There are many tasks in natural language processing that consist of associating
some kind of category to a group of words. Named Entity Extraction, Shallow
Parsing or Semantic Role Identification are three good examples. In this type
of tasks, we can identify two subtasks: one that finds the limits of the group of
words (chunk) and a second process that associates the correct tag to this group.
In this paper we present a series of experiments on a clear example of chunking:
the NER (Named Entity Recognition) task. We show how corpus transformation
and system combination techniques improve the performance in this task.
The NER task consists in the identification of the group of words that form a
named entity. IOB notation is usually employed to mark the entities in a corpus.
In this notation, the B tag denotes the beginning of a name, the I tag is assigned
to those words that are within (or at the end of) a name, and the O tag is
reserved for those words that do not belong to any named entity.
In the development of our experiments we have used the Spanish corpus dis-
tributed for the CoNLL’02 shared task [4], and the re-trainable tagger generator
TnT [1] based on Markov Models. In order to improve the performance of the
NER task we have increased the number of models. We have defined three trans-
formations that give us additional versions of the training corpus, and we have
trained TnT with them to obtain different taggers. The three transformations
improve the baseline given by TnT trained with the original corpus. Then we
have applied a stacking (machine learning) scheme to combine the results of the
models. The results of the experiments are presented in Table 1, we rely on the
well known recall, precision and Fβ=1 measures to compare the results.
References
1. Brants, T.: TnT. A statistical part-of-speech tagger. In Proc. of the 6th Applied
NLP Conference (ANLP00). USA (2000) 224–231
2. Carreras, X., L. Màrquez y L. Padró: Named Entity Extraction using AdaBoost. In
CoNLL02 Computational Natural Language Learning. Taiwan (2002) 167–170
3. Rössler, M.: Using Markov Models for Named Entity recognition in German news-
papers. In Proc. of the Workshop on Machine Learning Approaches in Computational
Linguistics. Italy (2002) 29–37
4. Tjong, E.F.: Introduction to the CoNLL’02 Shared Task: Language-Independent
Named Entity Recognition. In Proc. of CoNLL-2002. Taiwan (2002) 155–158
A Formally Verified Proof (in PVS) of the
Strong Completeness Theorem of Propositional
SLD–resolution?
Extended Abstract
resolvent(G, Cj ) resolvent(G, Cj )
n (via f )
i.h.
n-1 n-1(via f )
This lemma is not true if we consider the bodies of the clauses as
finite sets. Thus, we have generalized this lemma, by defining for each
program P a measure function hP and proving the following generalized
reduction lemma: if P ∪ {G} has a refutation, then for every atom A
of the body of G exists a clause C in P , such that its head is A, P ∪
{resolvent(G, C)} has a refutation and h P (resolvent(G, C)) < hP (G).
Roughly speaking, the function hP measures the “least effort” needed
to prove that an atom or a finite sets of atoms are logical consequences
of P .
In order to define this measure, we take into account that LC(P ) =
lf p(P ) and that lf p(P ) is built using the consequence operator T P , as
the least fixed point of the chain of consequences
TP0 (∅) ⊆ TP1 (∅) ⊆ . . . ⊆ TPk (∅) ⊆ . . .
First, we define the least exponent of an atom A with respect to a
program P . Intuitively, in the case that P |= A, this number represents
the first step in which A appears in the chain of consequences of P :
min{k : k < n ∧ A ∈ TPk (∅)}, if P |= A;
m exponentP (A) =
0, if P |6 = A.
Using it, we define the function hP as follows:
X
1 + min hP (B) : C ∈ CRP (A) , if P |= A;
hP (A) =
B∈body(C)
n
n , if P |6 = A.
where
References
1. K. R. Apt. From Logic Programming to Prolog. Prentice-Hall, Inc., Upper Saddle
River, NJ, 1996.
2. C. Baral. Knowledge Representation, Reasoning and Declarative Problem Solving
with Answer Sets. Cambridge University Press, 2003.
3. J. W. Lloyd. Logic Programming. Springer-Verlag, 2 edition, 1987.
4. A. Nerode and R. A. Shore. Logic for Applications. Springer-Verlag, New York,
NY, 1993.
5. S. Owre, J. Rushby, and S. N. PVS: A propotype of verification system. In
http://pvs.csl.sri.com.
6. R. Stärk. A direct proof for the completeness of SLD-resolution. In E. Börger, H. K.
Büning, and M. M. Richter, editors, Computer Science Logic, selected papers from
CSL ’89, pages 382–383. Springer-Verlag, Lecture Notes in Computer Science 440,
1990.
Fuzzy ontologies generation in GUMSe
Abstract
The main target of the ontology construction system here proposed is to solve one of
the main shortages in Internet search engines: “Its nonexistent capacity for semantic
deduction”.
As to endow these so-called semantic capabilities, the idea of automatically gener-
ating ontologies arises, although the majority of the attempts at it until now have failed
due to a basic lack: the definition of ‘processable conceptual relation’. Nowadays, the
main conceptual terms relation measurement is the co-occurrence of such terms as
Widyantoro [1] proposes, which is a main factor but neither determining nor unique.
The main innovation in this study is the redefinition of the idea of ‘processable
conceptual relation’ based on a new influential factor: “the physical distance between
terms”, and based on it, “the differentiation of kinds of conceptual relations”.
Two terms ti, tj , which are in the same sentence, are physically closer than being in
different paragraphs. The explanation of this idea dwells in the fact that the terms ap-
pearing in the same sentence are more bound to be strongly related by predicates with
the form “is a” or “is part of” or even become axioms of the ontology, than two terms in
different paragraphs although partially related to the same matter.
According to this theory, it is possible to find two kinds of different conceptual re-
lations, which does not mean they are independent. Such relations are known as physi-
cal relations and purely semantic relations.
The physical relations are mainly based on both summing up (“is part of”) and gen-
eralization (“is a”) relations since predicates with the form:
The car has wheels
Show in a simple and intuitive way the strong relation between the concept “car”
and the concept “wheel”. This example shows the great importance of the concept:
“physical distance”, for two concepts that are in the same sentence, might have a
strong degree of relationship. On the other hand, purely semantic relations prove that
even in a less intuitive way, some concepts may be related to other ones, as a text
talking about Spanish customs would show. The concept “Spain” would be related to
the concept “bullfighting” -even though Spain is not a bullfighter-, but in a less intui-
tive and therefore ‘weaker’ way. This could also apply to many other customs appear-
ing in different paragraphs, which shows once again that physical distance might be an
influential factor.
Logically, physical relations might also have an important semantic charge at the
same time. In a specific context it is possible to talk about the different types of car
wheels and its importance in rallies and so on, in this context, the relation between
“car” and “wheel” is even stronger, boosted by both physical and semantic relations
between them. Now, it appears the problem of delimitating how high is the degree of
matching between physical and purely semantic relations in two concepts. The relation
between two terms global coefficient is calculated by using the fuzzy intersection to
solve the overlapping zone between physical and purely semantic relation.
Acknowledg ments
Partially supported by PBC-03-004 PREDACOM project, JCCM, TIC2003-08807-C02-02
DIMOCLUST project, MCYT, and JCCM Cons. de Ciencia y Tecnología - European
Social Fund grants, Spain.
References
1. Widyantoro, D., Yen, J.: Incorporating fuzzy ontology of term relations in a search engine,
in: Proceedings of the BISC Int. Workshop on Fuzzy Logic and the Internet (2001) 155–
160.
2. WORDNET. A lexical database for the English language.
http://www.cogsci.princeton.edu/~wn/ (last visit 4 October 2004)
3. Olivas, J. A., Garcés, P., Romero, F. P.: An application of the FIS-CRM model to the
FISS metasearcher: Using fuzzy synonymy and fuzzy generality for representing concepts
in documents. Int. J. of Approximate Reasoning (Soft Computing in Recognition and
Search) 34 (2003) 201–219.
fun rec G =
partial_object_ext (carrier G)
(monoid_ext (fn x => fn y => mult G y x) (one G) Unity);
This seems quite natural, but stresses, in a very concrete way, Davenport’s
claim: if code is extracted directly from the statement it is clear that no trade-
off is acquired with respect to the logic in the proof. However, one could argue
that a “definitional” theorem should be proved in a “property-checking” way, so
obviously in a constructive manner. Being true, this observation does not capture
all the richness of our topic. In general, in CAT, properties are expressed in two
steps: first, a certain construction is proved computable, by constructing it as a
functional between two data types; second, this construction has the convenient
theoretical properties in order to ensure that the results obtained are the right
ones. Let us illustrate this rather speculative claim with another example from
monoids: there is a theorem constructing the product of two monoids. Then,
there is another theorem stating that this object has the right universal property.
It is not difficult (but not mandatory) to give a non-constructive proof of the
uniqueness part of the universal property (by contradiction, with explicit use of
the excluded middle principle). This example proves, in a toy case, that proof
extraction tools can be used in CAT to generate code, but that the underlying
logic could be non-constructive.
Our current research has produced a taxonomy of the lemmas already proved
in Isabelle [1], determining when this “extraction from statements” approach is
enough. More precisely, from the six auxiliary lemmas in Section 2 of [1], three
are directly expressed by constructive statements, and the other three can be
easily transformed into this kind of statement. The only result which seems
essentially different is the BPL itself. Its statement is the following (for the
necessary definitions and notations, we refer to [1]).
implies further difficulties from a constructive point of view. Some kind of in-
ductive definition would be necessary to deal with the series in a constructive
manner. Then, it should be possible to study if the Isabelle code extraction tool
can be applied in that inductive case.
Another research line is to analyze if the special characteristics of these state-
ments can help to give easier extraction algorithms, or, eventually, to establish
equivalence theorems with ML programs written independently.
References
1. J. Aransay, C. Ballarin and J. Rubio, Four approaches to automated reasoning with
differential algebraic structures, AISC 2004, LNAI vol. 3249, pp. 222-235.
2. S. Berghofer, Program Extraction in Simply-Typed Higher Order Logic, TYPES
2002, LNCS vol. 2646, pp. 21-38.
3. R. Brown, The twisted Eilenberg-Zilber theorem, Celebrazioni Arch. Secolo XX,
Simp. Top. (1967), pp. 34-37.
4. J. M. Davenport, Effective Mathematics: the Computer Algebra viewpoint, Lecture
Notes in Mathematics 873 (1981), pp. 31-43.
5. X. Dousson, F. Sergeraert and Y. Siret, The Kenzo program,
http://www-fourier.ujf-grenoble.fr/~sergerar/Kenzo/
6. M. Kaufmann, P. Manolios and J. Strother Moore. Computer-Aided Reasoning: An
Approach. Kluwer Academic Publishers, 2000.
7. Mathematics, Algorithms, Proofs, http://map.unican.es/
8. T. Nipkow, L. C. Paulson and M. Wenzel, Isabelle/HOL: A proof assistant for higher
order logic, Lecture Notes in Computer Science, 2283, 2002.
9. J. Rubio and F. Sergeraert, Constructive Algebraic Topology, Bulletin des Sciences
Mathématiques 126 (2002) 389-412.
Cross-Language Classification of Function Words
1 Introduction
1 X f (x, w, y)
OBL(w) = (1)
|(x, y) : ∃(x, w, y)| f (x, w, y) + f (x, y)
(x,y):∃(x,w,y)
1 X f (w, y)
R_HET ER(w) = (2)
|(y) : ∃(w, y)| f (w)
(y):∃(w,y)
1 X f (x, w)
L_HET ER(w) = (3)
|(x) : ∃(x, w)| f (w)
(x):∃(x,w)
For instance, the past participle “been” has a very homogeneous left context,
since it is followed by a very restricted set of verbal forms: “is”, “has”, “have”, . . .
Features OBL, R_HET ER, and L_HET ER are part of the 20 ones we use
to characterize the vector space of function words. Similarity between each pair
of vectors is computed by the Euclidean distance.
3 Classification of Function Words
PREP with_E
CONJ or_E
ART the_E
DEM this_E
AUX have_E
Then, we split each class in three monolingual word sets. It results in 15 basic
classes of function words (see Table 2).
Each basic class depicted in Table 2 will be used to classify other function
words within a particular language. For instance, the English classes will allow us
to classify rare and ambiguous function words. Such a monolingual classification
will follow a standard method based on context distribution.
ENGLISH
PREP with, at, under, for, from, by, on, as
CONJ or, and
ART the, a
DEM this, these, such, any
AUX have, has, is, are
PORTUGUESE
PREP para, por, com
CONJ e, ou
ART o, os, uma, um, as
DEM outros, qualquer, esta, este, esse, estas, outras, essa
AUX será, são, pode, foi
FRENCH
PREP avec, pour, sur, dans, par
CONJ et
ART la, le, les, l
DEM ces, cette, ce
AUX peut, est, a
Table 2. Monolingual classes of function words.
References
1. A Clark. Combining distributional and morphological information for part of speech
induction. In Proceedings of 10th European Chapter of the Association for Compu-
tational Linguistics, pages 59–66, 2003.
2. S. Finch and N. Chater. Bootstrapping syntactic categories. In 14th Annual Meeting
of the Cogntive Science Society, pages 820–825, 1992.
3. Pascale Fung. Compiling bilingual lexicon entries from a non-parallel english-chinese
corpus. In 14th Annual Meeting of the Cogntive Science Society, pages 173–183,
Boston, Massachusettes, 1995.
4. Pablo Gamallo, Joaquim Ferreira da Silva, and Gabriel P. Lopes. A divide-and-
conquer approach to acquire syntactic categories. In Yasubumi Sakakibara and
Georgios Paliouras, editors, Proceedings of the ICGI’04 Conference, Lecture Notes
in Computer Science. Berlin: Springer-Verlag., 2004.
5. H. Schutze. Part-of-speech induction from scratch. In ACL-93, Ohio State, 1993.
6. T.C. Smith and I.H. Witten. Probability-driven lexical classification: A corpus-based
approach. In PACLING-95, 1995.
Integrating syntactic information by means of
data fusion techniques ?
Extended abstract
In Information Retrieval (IR) systems, the correct representation of a document
through an accurate set of index terms is the basis for obtaining a good perfor-
mance. If we are not able to both extract and weight appropriately the terms
which capture the semantics of the text, this shortcoming will have an eect on
all the subsequent processing.
In this context, one of the major limitations we have to deal with is the
linguistic variation of natural languages, particularly those with more complex
morphologic and syntactic structures than English, as in the case of Spanish.
When managing this type of phenomena, the employment of Natural Language
Processing (NLP) techniques becomes feasible. Previous work has shown that lin-
guistic variation at word level can be eectively managed by taggers-lemmatizers,
improving the average precision attained by stemmers [5]. The next step con-
sists of applying phrase-level analysis techniques to reduce the syntactic variation
present in both documents and queries, identifying the syntactic structure of the
text by means of shallow parsing techniques that provide an eective balance
between descriptive power and computational cost [3]. However, the integration
of the information obtained by syntactic processing (usually in the form of com-
plex index terms) into the indices of an IR system is an open issue. As indicated
in [1], when simple and complex terms are used together, the assumption of term
independence is violated because words forming a complex term also occur in
the documents from which the syntactic dependency has been extracted.
In this article we propose to apply data fusion, a set of techniques for combin-
ing the results retrieved by dierent representations of queries and documents, or
by dierent retrieval techniques, to integrate index terms coming from dierent
text processing techniques. In particular, we consider the fusion of the following
indexing methods:
Single lemmatization. The lemmas of the content words {nouns, adjectives
and verbs, the grammatical categories which concentrate the semantics of a
? The research reported in this article has been partially supported by Min-
isterio de Educacion y Ciencia (MEC) and FEDER (TIN2004-07246-C02-01
and TIN2004-07246-C02-02), by MEC (HF2002-81), and by Xunta de Galicia
(PGIDIT02PXIB30501PR, PGIDIT02SIN01E and PGIDIT03SIN30501PR)
Fig. 1. Weighted mixing of index terms.
text{ are used as index terms. For lemmatizing the documents, we employ
MrTagoo [7], a high-performance part-of-speech tagger based on a second
order Hidden Markov Model, with lemmatization capability, whose input is
provided by a linguistically-motivated preprocessor module [8].
Syntactic dependency pairs. This retrieval method is based on complex
terms derived from the syntactic dependencies existent in the documents
and in the queries. We employ a shallow parser based on a cascade of nite-
state transducers consisting of ve layers [3], whose input is the output of
the tagger-lemmatizer. The result we obtain is a list of syntactic dependency
pairs between the heads of the constituents identied from the text.
Syntactic dependency pairs ltered by relevance feed-back. The index-
ing process is the same described in the previous case, but the querying
process is performed in two stages:
1. The lemmatized query is submitted to the system.
2. The n top documents retrieved by this initial query are employed to
select the most informative dependencies, which are used to expand the
query in order to obtain the nal set of documents. These dependencies
are selected automatically using Rocchio's approach [9] to feedback.
Locality-based retrieval. In this case, we apply a pseudo-syntactic ap-
proach [4] based on the proximity between index terms.We have adapted
the original proposal of Kretser and Moat [6], based on a similarity mea-
sure computed as a function of the distance between terms {in our case
single lemmas{ where each occurrence of a query term has an in
uence on
the surrounding terms.
With regard to the combination of the results oered by these retrieval methods,
the following data fusion techniques have been considered:
Weighted mixing of index terms. This fusion method, illustrated in Fig. 1,
merges the sets of index terms extracted from the documents by the dierent
text processing techniques and assigns to each of them its own weights to
be used within the Smart vectorial IR engine [2], which will take them into
account when computing vector distances during the retrieval phase.
In our experiments we have performed exhaustive tests, trying several
weights for each indexing method and also using dierent formulae to com-
pute vector components. The best results were obtained when single terms,
Fig. 2. Reindexing retrieved documents.
extracted by lemmatization, were assigned weights higher than the ones as-
signed to complex index terms.The main problem with this approach is how
to nd the right set of weights, capable of oering a uniform behaviour with
all of the queries being considered, because weights with good results in some
queries do not necessarily maintain their performance with other queries.
Reindexing retrieved documents. In this second method we propose a two-
step approach. As shown in Fig. 2, we perform a rst retrieval process using
single terms extracted by lemmatization. Then, we select the top k ranked
documents and reindex them using complex terms based on dependency
pairs. The nal list of documents is built by reordering the top documents
retrieved in the rst step according to the results of the second phase.
The idea behind this approach is to perform a more precise processing on
the most promising documents and use it to improve the nal ranking. We
have tested this proposal by reindenxing dierent amounts of top ranked
documents and using dierent methods to extract dependency pairs for the
reindexing step. The results obtained were not as good as was expected. The
main reason for this poor behaviour was the high intercorrelation shown by
the index terms extracted during the second step, where the classical formu-
lae to compute vector components does not seen to oer good performances.
List merging based on rank positions and mutual precision. The last
method considered in this paper tries, in contrast with the previous ones,
to perform a fusion of results adapted to each individual query. What
this proposal does is to merge lists of retrieved documents obtained using
dierent indexing methods. As shown in Fig. g-mixing, the nal ranked
list takes into account both the ranking position of each document in the
lists that retrieved it and a value that measures the condence that each
list assigns to those positions.
These condence measures are computed independently for each query and
for each list of retrieved documents, using the precision values obtained for
dierent subsets of the list, taking a combination of the other document lists
as a reference to compute them. As a consequence of this way of computing
Fig. 3. Mutual precision based merging.
these condence measures, the main drawback of this method is that it
improves the ranking position of both relevant and non-relevant documents
that are common to dierent lists. So, this approach can be expected to oer
promising results when dealing with high precision lists of documents.
The corpus employed in our evaluation is the Spanish monolingual corpus
from CLEF (Cross Language Evaluation Forum ), composed by 215,738 news
reports. The set of topics consists of 100 queries, formed by three elds: a brief
title, a one-sentence description, and a more complex narrative specifying the
relevance criteria. All these three elds have been employed to build the running
queries in our experiments, giving a dierent relevance to each of them.
References
1. M. Narita and Y. Ogawa. The use of phrases from query texts in information
retrieval. In ACM SIGIR 2000, pages 318{320, Athens, Greece, 2000.
2. C. Buckley. Implementation of the SMART information retrieval system. Technical
report, Department of Computer Science, Cornell University, 1985.
3. J. Vilares and M. A. Alonso. A grammatical approach to the extraction of index
terms. In Proc. of RANLP 2003, pages 500{504, Borovets, Bulgaria, Sept. 2003.
4. J. Vilares and M. A. Alonso. Dealing with syntactic variation through a locality-
based approach. Lecture Notes in Computer Science (LNCS), 3246:??{??, 2004.
5. J. Vilares, M. A. Alonso, and F. J. Ribadas. COLE experiments at CLEF 2003
Spanish monolingual track. In Working Notes for the CLEF 2003 Workshop, pages
197{206, Trondheim, Norway, Sept. 2003.
6. O. de Kretser and A. Moat. Locality-based information retrieval. In Proc. of 10th
Australasian Database Conference (ADC'99), pages 177-188, Singapore, 1999.
7. J. Grana, M. A. Alonso, and M Vilares. A common solution for tokenization and
part-of-speech tagging: One-pass Viterbi algorithm vs. iterative approaches. in Lec-
ture Notes in Computer Science (LNCS), 2448:310, 2002.
8. F. M. Barcala, J. Vilares, M. A. Alonso, J. Grana, and M. Vilares. Tokenization
and proper noun recognition for information retrieval. In DEXA Workshop 2002,
pages 246250. IEEE Computer Society Press, 2002.
9. J. J. Rocchio. Relevance Feedback in Information Retrieval. In G. Salton, editor,
The SMART Retrieval System{Experiments in Automatic Document Processing.
Prentice Hall, Englewood Clis, NJ, 1971
!"#%$'&() #*)+,.-/0
1 ,)+23)45& ) 67
0%'& ,89&:;,-/0=<>?)&A@"BC%& 34D
E
F0GIHKJMLONQPQROSGOTVUXWZYM[INW\R^]_Ga`cbedXWMROfIGIFcdXLOgihjhedOR^kcG lBGnmoJMprqId\stsuJIRvkG HKJ\Lvwud\xOyQP
z*{}|~u~j
a/|(
~Q(v(c/ZXuX| nt¡£{*¢Z¤vCZo¥Qv¦\~a¡u§X{¨}~|rCBX~(MM\vQ~
1 Introduction
Data mining is an essential step in the process of knowledge discovery in da-
tabases that consists of applying data analysis and discovery algorithms that
produce a particular enumeration of structures over the data. There are two
types of structures: models and patterns. So, we can talk about local and global
methods in data mining. In the case of local methods, the simplest case of pattern
discovery is finding association rules [1]. The initial motivation for association
rules was to aid in the analysis of large transactional databases. The discovery
of association rules can potentially aid decision making within organizations.
Another approach is integrating the data mining process into the development
of Knowledge Based Systems.
Since the problem of mining association rules was introduced by Agrawal
in [1], a large amount of work has been done in several directions, including
improvement of the Apriori algorithm, mining generalized, multi-level, or quan-
titative association rules, mining weighted association rules, fuzzy association
rules mining, constraint-based rule mining, efficient long patterns mining, main-
tenance of the discovered association rules, etc. We want to point out the work
in which a new type of association rules was introduced, the inter-transaction
association rules [3]. Temporal data mining can be viewed as an extension of
this work.
Temporal data mining can be defined as the activity of looking for inte-
resting correlations or patterns in large sets of temporal data accumulated for
other purposes. It has the capability of mining activity, inferring associations of
contextual and temporal proximity, some of which may also indicate a cause-
effect association. This important kind of knowledge can be overlooked when the
temporal component is ignored or treated as a simple numeric attribute [4].
We can find in the literature a large quantity of temporal data mining tech-
niques. We want to highlight some of the most representative ones. So, we can
talk about sequential pattern mining, episodes in event sequences, temporal as-
sociation rules mining, discovering calendar-based temporal association rules,
patterns with multiple granularities mining, and cyclic association rules mi-
ning. However, there is an important form of temporal associations which are
useful but could not be discovered with this techniques. These are the inter-
transaction associations presented in [3]. The introduction of this type of asso-
ciations was motivated by the observation that many real-world associations
happen under certain context, such as time, place, etc. In the case of temporal
context, inter-transactional associations represents associations amongst items
along the dimension of time. Due to the number of potential association be-
comes extremely large, the mining of inter-transaction association poses more
challenges on efficient processing than classical approaches. In order to make the
mining of inter-transaction associations practical and computationally tractable,
several methods have been proposed in [5].
In order to reduce the time execution of the algorithm, in this paper we pro-
pose an extension of T SET which is based in the reusing of structures generated
by previous data mining process.
2 An Iterative Mining Process
If the user sets the minsup value to 1, instead of generating the whole struc-
ture starting from the scratch, the algorithm iterates over the structure looking
for nodes with support less than or equal to the chosen value. In Figure 2 we
can see the generated structure.
To check the usefulness of this approach, an empirical evaluation is being
carried out over several kind of datasets. In particular, we are studying the algo-
rithm with real databases which contain phytosanitary treatment information.
Fig. 2. The second extended set-enumeration tree structure
References
1. R. Agrawal, T. Imielinski, and A. N. Swami. Mining association rules between sets
of items in large databases. In P. Buneman and S. Jajodia, editors, Proc. of the
ACM SIGMOD Int. Conf. on Management of Data, Washington, D.C., May 26-28,
1993, pages 207–216. ACM Press, 1993.
2. F. Guil, A. Bosch, and R. Marı́n. TSET: An algorithm for mining frequent temporal
patterns. In Proc. of the First Int. Workshop on Knowledge Discovery in Data
Streams, in conjunction with ECML/PKDD 2004, pages 65–74, 2004.
3. H. Lu, L. Feng, and J. Han. Beyond intra-transaction association analysis: Mining
multi-dimensional inter-transaction association rules. ACM Transactions on Infor-
mation Systems (TOIS), 18(4):423–454, 2000.
4. J. F. Roddick and M. Spiliopoulou. A survey of temporal knowledge discovery
paradigms and methods. IEEE Transactions on Knowledge and Data Engineering,
14(4):750–767, 2002.
5. A. K. H. Tung, H. Lu, J. Han, and L. Feng. Efficient mining of intertransaction asso-
ciation rules. IEEE Transactions on Knowledge and Data Engineering, 15(1):43–56,
2003.
6. Zhi Hua Zhou. Three perspectives of data mining (book review). Artificial Intelli-
gence, 143:139–146, 2003.
Data Mining with Scatter Search?
I.J. Garcı́a del Amo, F. Garcı́a López, M. Garcı́a Torres, B. Melián Batista,
J.A. Moreno Pérez, J.M. Moreno Vega and Raquel Rivero Martn
1 Introduction
Processing Intelligent Information requires efficient tools to extract the useful in-
formation stored in databases. Data Mining and Knowledge Discovery are power-
ful technologies for the extraction of information from large databases. Heuristic
approaches are becoming very useful for designing efficient tools in Data Mining
[1]. Most of data mining tasks are solved by the application of Machine Learning
techniques. Three central paradigms for the application of Machine Learning [9]
in Data Mining are Clustering, Instance-Based classification and Feature Selec-
tion. We describe the main components of the Scatter Search Metaheuristic and
their specific design to solve standard problems in these tasks. This method has
been tested for these or similar problems ([3], [2]).
Scatter Search [8] is an evolutionary population-based metaheuristic that
constructs solutions by combining others in an evolving set of solutions, named
reference set (Ref Set). The two main differences between Scatter Search and
other classical population-based evolutionary procedures in Data Mining [7] like
Genetic Algorithms [6] are the size of the evolving set of solutions and the way
the method combines the existing solutions to provide new ones. The evolving
set Ref Set has relatively small or moderate size (typical sizes are 10 or 15, see
[8]). Scatter Search combines good solutions to construct others exploiting the
knowledge of the problem at hand in an intelligent way. Genetic Algorithms are
also evolutionary algorithms in which a population of solutions evolves by using
the mutation and crossover operators. The mutation and crossover operators
used in Genetic Algorithms have a significant reliance on randomization to create
new solutions.
In a Scatter Search algorithm [8], a moderate sized set of solutions, the reference
set Ref Set, evolves due to mechanisms of intelligent combination of solutions.
This RefSet is generated and iteratively updated attempting to intensify and
diversify the search. After intelligently combining the solutions in the reference
set, a local search procedure is applied to improve the resulting solution, and
?
This research has been partially supported by the Ministerio de Ciencia y Teconologı́a
through the project TIC2002-04242-C03-01; 70% of which are FEDER funds.
the RefSet is updated to incorporate both good and disperse solutions. These
steps are repeated until a stopping condition is met. The method provides not
only a single heuristic solutions, like other metaheuristics, but a reduced set of
disperse high quality solutions.
The principles of the Scatter Search metaheuristic were first introduced in the
1970s as an extension of formulations for combining decision rules and problem
constraints. This initial proposal generates solutions taking account of charac-
teristics in several parts of the solution space [4]. Scatter Search includes five
main components processes: 1) Diversification Generation Method, that gener-
ates a set of diverse solutions, 2) Improvement Method, that improves a solution
to reach a better solution, 3) Reference Set Update Method, which builds and
updates the reference set consisting of Ref SetSize good solutions, 4) Subset
Generation Method, to produce subsets of solutions of the reference set, and 5)
Solution Combination Method, that combines the solutions in the produced sub-
sets. A comprehensive description of the fundamentals of Scatter Search can be
found in [5].
References
1. H.A. Abbass, C.S. Newton , R. Sarker. Data Mining: A heuristic Approach. Idea
Group (2002).
2. F. Garcı́a López, M. Garcı́a Torres, B. Melián Batista, J.A. Moreno Pérez and
J.M. Moreno Vega. Solving Feature Subset Selection Problem by a Parallel Scatter
Search, European Journal of Operational Research, 2005, to appear.
3. F. Garcı́a López, B. Melián Batista, J.A. Moreno Pérez and J.M. Moreno Vega.
Parallelization of the Scatter Search for the p-median problem, Parallel Computing,
29 (2003) 575-589.
4. Glover, F., Heuristics for Integer Programming using Surrogate Constraints, Deci-
sion Sciences 8, (1977) 156–166
5. Glover, F., Laguna, M., Martı́, R. Fundamentals of Scatter Search and Path Relink-
ing Control and Cybernetics, 39, (2000) 653-684
6. Goldberg, D.E., Genetic Algorithms in Search, Optimization and Machine Learning,
Addison Wesley, (1989)
7. Ghosh, A. and Jain, L.C. (Eds.) Evolutionary Computation in Data Mining Studies
in Fuzziness and Soft Computing, 163. Springer (2005)
8. Laguna, M. and R. Martı́, Scatter Search: Metodology and Implementations in C,
Kluwer Academic Press, (2003).
9. Mitchell, T. Machine Learning, Series in Computer Science, McGraw-Hill, (1997).
This article was processed using the LATEX macro package with LLNCS style
Web Usage Mining Project for Improving Web-based
Learning Sites
Extended Abstract
Relationships with the users are paramount when trying to competitively develop ac-
tivities in any web environment. Information collected and stored by Web servers is
the main source of data for analyzing user navigation patterns. A successful e-CRM
depends mainly on the capabilities of the organization to identify and classify users in
order to model and predict her/his behaviour on the site. Web site personalization is
one of the keys to do it [2]. Different web mining approaches and techniques have
been proposed in order to improve the sites by creating adaptive Web sites, based
mainly on the analysis of user access logs [5].
Higher education (HE) is an example of an activity that is having success when be-
ing developed through the web. Students and professor are provided with easy to use
tools and are not subject to any temporal constraint. However, it is very difficult and
time consuming for educators to thoroughly track and assess all the activities per-
formed by all learners on all these tools. Moreover, it is hard to evaluate the structure
of the course content and its effectiveness on the learning process [8].
E-learning environments lack a closer student-educator relationship. The lack of
this relation is manifested in facts such as: teacher does not really control the evolu-
tion of his students, and students cannot express their problems and deficiencies in a
natural way.
The roots of this problem lays in web-based learning environments, such as Vir-
tual-U [3] and WebCT [6], including course content delivery tools, synchronous and
asynchronous conferencing systems, polling and quiz modules, virtual workspaces for
sharing resources, white boards and other tracking tools but not including specific
tools to track and analyze student behaviour and their evolution.
In fact, these environments provide the educator with access summary information
such as more visited pages, favourite communication method, and other statistics.
Nevertheless this information is not enough to analyze the behaviour of each student
when the number of students increases.
Although data mining has been successfully applied in e-commerce sites to im-
prove the lost 1-to-1 relationship, it is not yet engaged into e-learning systems. Never-
theless, web-based learning systems can profit from the integration of data mining and
the e-learning technology [4].
The present paper presents advances of an e-learning project in which data mining
techniques are applied to improve the relationship between professor and student. Ac-
cordingly, typologies of students will be obtained based not only on their navigation
and demographics but also on behaviour evolution.
Thus, five steps will be needed:
1. Data pre-processing: clean and prepare the Apache log file and load the clean data
into a relational database
2. OLAP analysis: design a multidimensional structure in which the main factors un-
der analysis: sessions, pages, user data, user behaviour (contents and navegation)
will be taken as dimensions and later build OLAP cubes in order to analyse the re-
corded data
3. Pattern discovery: application of data mining algorithms. Firstly descriptive algo-
rithms (clustering and association) will be applied to obtain typologies of users.
Then classification techniques will be used in a later step to classify behaviours ac-
cording to historical patterns
4. Pattern evaluation: All the patterns obtained will be evaluated to distinguish the
patterns that really help to better achieve the site goals.
5. Recommendation engine: integrate the discovered patterns in the online component
to provide personalized environment to learners.
Consequently, results of the project will make it possible to design improved e-
learning tools mainly in two aspects: on the educator side, their tasks will be devel-
oped in an informed way, ie, he will have measures to better follow the learning proc-
ess and to analyze the web course structure effectiveness; on the students side, they
will obtain a personalized environment that in a next future, will recommend them ac-
tivities and/or resources that would favour and improve their learning. An added value
of these tools will be that the site administrator will have parameters to improve the
site efficiency and adapt it to the behaviour of its users.
The development of this project will highlight important results, both methodo-
logical and application. A case study to show the obtained results will be presented.
The case study will be performed making use of the data from Universidad de Can-
tabria, and WebCT tool will be used. The project will be performed following the
CRISP-DM [1] process model. As data mining software mainly WEKA [7] will be
used, while Microsoft SQL Server will be used as the data repository.
References
1. CRISP-DM. http://www.crisp-dm.org
2. Dai, H., Mobasher, B. A road map to more effective web personalization: Integrating
domain knowledge with web usage mining. In Proc.of the International Conference on
Internet Computing 2003 (IC’03), Las Vegas, Nevada, June 2003.
3. Groeneboer, C, Stockley, D. Virtual-u: A collaborative model for online learning environ-
ments. 2nd Int Conference on Computer Support for Collaborative Learning, Toronto, Can-
ada, December 1997.
4. Hanna, M. Data Mining in the e-learning domain. Campus Wide Information Systems. 16
Jaunary 2004. vol. 21 no. 1 pp. 29-34(6). Research article ISSN: 1065-0741
5. Perkowitz, M., Etzioni, O. Adaptive web sites: Automatically synthesizing web
pages. In AAAI/IAAI, pages 727–732, 1998.
6. WebCT, http://www.webct.com/
7. WEKA http://www.cs.waikato.ac.nz/ml/weka/
8. Zaïane, O. Web Usage Mining for a Better Web-Based Learning Environment. Proc. of Con-
ference on Advantage Technology for Education. Alberta, Canada. 2001.
Similarity Queries in Data Bases
Using Metric Distances -
From Modeling Semantics to its Maintenance
Extended Abstract
Similarity search has emerged and become a fundamental requirement, even for data
base applications and data base management systems. A theoretical framework for
similarity queries is given in [1]. Unfortunately similarity search based on the attrib-
ute values itself does not fit all user needs. For example in the case of a tourism in-
formation system the attribute values “San Augustin” and “Playa del Ingles” should
be identified as similar because both are villages in the southern part of Gran Canaria,
even if their attribute values are anything but similar. Additional concepts to model
and maintain semantic background information are needed to support such enhanced
similarity queries.
There are several approaches to represent similarities which can’t be derived di-
rectly from attribute values. One of them is to use points in multidimensional spaces
as substitutes. Each attribute domain can have its own space. Attribute values are
represented as points in that space and the distances correspond reciprocally to their
similarity. These attribute similarities build the basis for computing whole objects
similarities.
At the institute for applied knowledge processing a prototype for such a system has
been implemented and tested. We called it VQS – Vague Query System [2]. One of
the main goals of this system was that it can be used as an add-on to still existing data
bases. Therefore we concentrated on relational and object relational data bases. The
semantic background information (the mapping from attribute values to multidimen-
sional points) lies in Numeric Coordinate Representation Tables (NCR-Tables).
These build quasi the similarity base, comparable to a knowledge base. An extra
query language (VQL – vague query language) allows formulating similarity queries.
It looks like SQL but, supports a new operator “is” which means “is similar to”. Any
combination of similarity conditions with non similarity conditions and, even similar-
ity joins are supported by this system. A practical test using the database of a large
tourism information system showed that the system works quite well.
One crucial question is left: “How to build an adequate representation of attribute
values in form of multidimensional points?” In some cases this mapping is obvious.
For localities we can use the geographical coordinates, for colors a three dimensional
representation (red, blue, green) could be chosen, and so on. For all other cases,
where such a simple mapping can not be found the NCR-table has to be maintained
manually. Newer research at our institute showed that an automatic finding of an
appropriate NCR-representation is just possible when all similarity values between
each pair of attribute values are known [3]. Otherwise – which is the case in practice
– for the number of dimensions needed just an upper bound can be given and the
finding of multidimensional points can be modeled as an optimization problem.
Therefore we are now working on using learning techniques such as reinforcement
learning for adjusting the similarity base step by step.
References
1. Jagadish, H. V., Menelzon A. O., Milo T.: Similarity-based Queries. Proceedings of the
fourteenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database sys-
tems, ACM, 1995
2. Küng, J., Palkoska, J.: VQS – A Vague Query System Prototype. Proc. Eighth International
Workshop on Database and Expert Systems Applications, September 1-2, 1997, Toulouse,
France, IEEE Computer Society Press (1997) 614–618
3. Csaji, B.C., Küng, J., Palkoska J., Wagner R.: On the Automation of Similarity Information
Maintenance in Flexible Query Answering Systems. Proceedings of the 15th International
Conference on Database and Expert Systems Applications - DEXA 2004, Springer, Lecture
Notes in Computer Science (2004) 130-140
A WEB-CASE Tool Prototype For Hybrid
Software Development
1 Introduction
When we are developing software, based or not based on knowledge, develo-
pers choose a methodology depending of the problem, the prior knowledge, the
available tools or other factors. These methodologies lead us to the production
process determining the documents and artifacts we must generate, the activities
to do and the order that we must follow.
The main artifact generated during the software development process, is a
group of models [1]. The models of rst stages in the process are often specied
using natural language. However, this technique has problems like ambiguity, un-
certainly or imprecision. Several alternatives have proposed a solution by means
of the utilization of a structured natural language, description language, des-
cription languages of design, graphic notation or formal specications, that are
similar to the successfully used in the nal stages of software development [4].
We propose to use these techniques in the organization modelling and requi-
site level scope, combining the forms and diagrams to model complex systems
and a CASE tool for managing these artifacts. But, in a lot of cases, the models
and documentation that we generate and maintain is very complex. So, we need
software tools that guide us in the generation of these artifacts, its maintenance,
and the control of the integrity of changes.
References
1. Object management group. In MDA Guide Version 1.0.1. OMG Document:
omg/2003-06-01, 2003.
2. J. J. Cañadas, A. Bosch I. M. del Águila, and S. Túnez. An intelligent system
for therapy control in a distributed organization. In Proceedings of Information and
Communication Technology (EurAsia-ICT 2002), Springer-Verlag. Lecture Notes in
Computer Science, 2510:1926, 2002.
3. I. M. del Águila, J. J. Cañadas, J. T. Palma, and S. Túnez. Integration of develop-
ment methodologies for the building of knowledge intensive multiagent systems. In
In Proc. of the First Int. Conf. on Integration of Knowledge Intensive Multi-Agents
Systems, KIMAS'03, pages 203208. IEEE Press, 2003.
4. G. Schreiber, H. Akkermans, A. Anjewierden, R. deHoog, N. Shadbolt, W. Van-
deVelde, and B. Wielinga. Knowledge Engineering and Management: The Com-
monKADS Methodology. The MIT Press, 1999.
An Augmentative Communication System based on
Adaptive Evolutionary Hypermedia Systems
1 Introduction
The elements that this approach manages in order to design adaptive evolutionary
communication systems are:
• Ontology model, which represents each conceptual world and describes a
specific knowledge domain by means of concepts and relationships between them
(a semantic net) [4]. The ontology, in this case, provides the link between the
symbols used in the communicator (image, sound or text adapted to each
person’s knowledge) and the entities of the real world they represent. In addition,
ontologies both allow a global vision of each conceptual world based on the
concepts employed by the users, and allow to standardize into a single model the
knowledge about the people with the communicative difficulties, the knowledge
about the people around them (i.e. relatives, tutors, rehabilitators, etc.), and the
media used for communication (i.e. templates, agendas, etc.).
• User model, which provides knowledge about the people with communication
disabilities and consists on the following components:
o Knowledge domain represented by means of an ontology. In addition, the
educators can define different semantic views of the whole knowledge
domain taking into account the user profile. For example, figure 1 shows a
partial view of an ontology that describes a scenario of shopping.
The process followed by the tutor to create a new view consists in hiding
in the semantic net the concepts, items and conceptual relations he
considers not related to the current view. The system automatically keeps
the integrity of the view, for example, if after hiding a relation some
concepts get disconnected, they are also automatically hidden. In the same
way, when a concept is hidden all the items associated to it and all the
relations starting or arriving to the concept are also hidden.
Once the semantic net is built, the system will be able to automatically
generate the interaction templates from it. These templates constitute one
of the possible media used to establish the communication with the user
and are represented in hypermedia format.
o User profile, containing the particular characteristics of users
(communication habits, training or therapy objectives, and interaction
preferences). To construct and select it the user (and relatives or educators)
must provide this information. This artefact allows the adaptation and
selection of the semantic view that best fits to a particular user.
• Adaptation by feedback. By means of a transition matrix, the system captures
the conceptual relationships followed by the person during navigation. The
matrix has one row and one column for every concept represented in the
knowledge domain. Based on the analysis of the matrix, the system can identify
new conceptual relationships that are needed or existing conceptual relations that
should be removed in the knowledge domain. In addition, this mechanism helps
to detect progress or problems in the evolution of the patients.
3 Architecture of Adaptive Evolutionary Communication System
The architecture of our adaptive model has two tiers: the system (communicator) and
the meta-system [6]. This division allows us to separate on the one hand the
interaction, communication and user adaptation and on the other hand its evolution
process.
The communicator is used by the user and the meta-system by the educators. The
cognitive, interaction, design and adaptation aspects have to be differentiated to avoid
the coupling. Thereby, evolution/adaptability can be done more easily and safely
because the architecture components are independent.
The system is itself divided in four parts, which allows us to tackle separately the
aspects of knowledge representation (memorization system), construction of partial
views focused to one o several concrete knowledge subdomains (presentation
system), navigation of the built views (navigation system), and personalization of the
navigation process (learning system) [2].
The structure offered to navigate is different depending on the used platform. If a
personal computer is used the interface has two frames: at the left the semantic net
and at the right some information about the element selected in the net. It provides a
semantic and contextual mode of navigation. If it has not been predefined by the
educator, the user can choose the navigation mode that he desires. Depending of the
chosen mode (free navigation, conceptual navigation, navigation by conceptual
relation or navigation by prerequisites), some options in the semantic net are disabled
according to the user features and his interests. So, the adaptation is carried out
depending on the user, but also on the navigation mode used in that moment.
In other case, when a PDA or other embedded system is used, the interface is based
in templates. Every template only shows the available options in the current moment,
without additional information. Each option is represented by means of an icon which
can also have associated text and audio.
Figure 2 shows a possible use of the PDA in a scenario based on figure 1. As it can
be seen, in some templates all the options must be selected in a particular sequence,
while in others it is enough that the user selects one or several options in any order.
Fig. 2. Shopping with the PDA
References
GREYC, Team ISLanD, Computer Science Laboratory, University of Caen, F14032, Caen
Cedex, France
{troy, beust}@info.unicaen.fr
1 Introduction
The quantity of information exchanged every day on the Web doesn’t stop increasing.
Generalist search engines, like Google or Yahoo, return hundreds of links from a
user’s request. These links are ordered in lists, but such lists do not allow users to
have a global view on returned results. However, such view would reveal the
principal categories of information from returned Web pages. In order to give users a
global view on documents’ sets, it can be interesting to use graphical representations
[3]. A lot of NLP tools represent sets of documents in 2 or 3-Dimensional space (we
can cite [4], [5], [6], [7], for such works), each one uses specific visualization
methods.
The visualization method used in ProxiDocs is calling “cartography”. A map of a
set of documents reveals proximities and links between textual entities (like words,
documents, etc.); we can compare this visualization method with road maps which
reveals proximities and ways between towns. Since 2001, the two metasearch engines
KartOO [8] and MapStan [9] are available on the Web2. These tools return a map
ProxiDocs returns thematic maps from a set of documents (for example: a corpus of
newspaper's articles or a set of Web pages returned by a search engine) and themes
specified by the user. A theme will be defined by a set of words chosen by the user.
For example, the theme “Sport” could be specified by the following terms: “winner”,
“looser”, “tennis”, “football”, “Zinedine Zidane”, etc. We advise the ThemeEditor3
tool [10] develop in our research team for a such task of themes contruction’s.
The first stage consists in counting words of each theme in each document. So, we
associate a numbers’ list with each document, and, each list constitutes a N-
Dimensional vector (N is the number of themes specified by the user). For instance, if
a document contains 3 words of the theme “Economy”, 2 words of the theme “Sport”
and 7 words of the theme “Education”, the vector associated to this document is (3, 2,
7).
The next stage consists to project these N-Dimensional vectors on a 2-Dimensional
space. In order to realize a such projection, we use a method called Principal
Components Analysis [11]. Each document is then represented by a point in a 2-
Dimensional space, so, first maps can be constructed. On these maps, proximities
between different points could inform the user on thematic similarities between
different documents.
In order to underline documents’ subsets with similar themes, we use a clustering
method called Ascendant Hierarchical Clustering [11]. Thus, thematic maps reveal
clusters of documents, each cluster containing documents dealing with similar
subjects.
4 Conclusion
Proxidocs allows its users to visually identify main thematic tendencies of a set of
documents. We work now on the application of thematic cartography to results of
search engines in order to distinguish principal thematic categories of information
contain in returned Web pages. In an technological watch context, our tool could also
reveal thematic changes of a set of documents regarding to the time. In order to
improve our tool, we have to measure the “quality” of thematic maps returned to
users. This assessment will be difficult because we will have to elaborate a protocol
taking account of the point of view of each user on themes used during the
construction of thematic maps.
References
1 Introduction
The multi-tier content selection process to be employed in this paper involves a do-
main specific characterization, a set of keywords, a relevance feedback tier, and an
automatic categorization algorithm.
The first tier of selection corresponds to a domain specific given classification of
the documents - for instance, in a digital newspaper, the assignment of news items to
newspaper sections: national, international, sport, etc. The user must assign a weight
to them (su).
For the second tier, the user enters a set of keywords - with an associated weight -
to characterise his preferences: appearance of these keywords in the documents indi-
* This research has been partially funded by the Spanish Ministerio de Ciencia y Tecnología
(TIC2002-01961)
cates a possible interest. These keywords are stored, for each user u, as a term weight
vector (ku).
For the third tier the user must choose - and assign a weight to them - a subset of
those categories in the first level of Yahoo! [4]. This information is stored as a matrix
where rows correspond to categories and columns correspond to users (Ggu). These
categories are represented as term weight vectors (g) by training from the very brief
descriptions of the 14 first level of Yahoo! Spain categories entries [4].
In the fourth tier, short term interests [1, 5] are represented by means of feedback
terms obtained from user provided feedback over the documents he receives [2]. The
term weight vector for each user (tu) is taken to represent the short term interests of
that user. These correspond to information needs whose interest to the user wanes
after a short period of time, so their weight must be progressively decreased over
time.
Documents are downloaded from the web as HTML documents. For each docu-
ment, title, category, URL and text are extracted and stored for subsequent process-
ing. Term weight vector representations (dd) are obtained by application of stop lists,
stemmer, and the tf · idf formula for computing actual weights [6].
As each web document has a preassigned section, it is assigned the weight associated
with the corresponding section in the particular user model. The similarity between a
document and a category is computed using the cosine formula of the vector space
model [6, 7]. The similarity between a document and the keywords of a user model is
computed also using the cosine formula.
⎛ 14 ⎞
⎜ ∑ Giu sim(d d , gi ) ⎟
α su + β ⎜ i =1 14 ⎟ + χ sim(d d , ku ) + δ sim(d d , tu )
⎜
⎜
⎝
∑
i =1
Giu ⎟⎟
⎠
(1)
rdu =
α + β + χ +δ
The short term model is obtained as a result of the process of adaptation of the user
model. The user receives a web document that contains an automatically generated
personalised summary [3] for each of the 10 web documents that the system has
found more relevant according to his user profile. With respect to this information the
user may give positive or negative feedback. The feedback terms of the short term
model are obtained from the news items for which feedback has been provided [2].
The similarity between a document and the feedback keywords of a user model is
computed also using the cosine formula. When all documents have been ordered with
respect to the various reference frameworks, the results are integrated using a particu-
lar combination of them.
4 Evaluation
Experiments are evaluated over data collected for 106 users and the news items corre-
sponding to three weeks – the 14 working days of the period 1st-19th Dec 2003 - of
the digital edition of the ABC Spanish newspaper. The average of news items per day
is 78.5.
To carry out the evaluation, judgements from the user are required as to which
news items are relevant or not for each of the days of the experiment. To obtain these
judgements users were requested to check the complete set of news items for each
day, stating for each one whether it was considered interesting (positive feedback),
not interesting (negative feedback) or indifferent (no feedback). Users were explicitly
asked not to confine their judgements on interest to relevance with respect to the
initial user profiles they had constructed on first accessing the system, but rather to
include any news items that they found interesting on discovery, regardless of their
similarity with respect to their initial description of their interest. It is hoped that
enough information to cover these rogue items will be captured automatically and
progressively by the system through the feedback adaptation process.
Since our experimental set up combines a binary relevance judgement (indifferent
is considered as not interesting ) from the users and a ranking of news items provided
by the system, it was decided to use normalised recall and precision [6] as our evalua-
tion metrics. In addition, with respect to equal relevance values for consecutive posi-
tions of the ranking, the average ranking of the whole set of conflicting positions has
been taken as ranking for each and all of them. This adjustment avoids the problem of
ordering items at random within the ranking when they have equal relevance.
The following experiments have been carried out to check the validity of the pro-
posed model for the selection process. Each experiment combines different possibili-
ties for long term modeling - only sections, L(S), only categories, L(C), sections and
categories, L(SC), sections and keywords, L(SK), categories and keywords, L(CK),
sections, categories and keywords, L(SCK). This implies giving different values to
the parameters α, β, χ and δ of formula (1). For example, L(SC) is the combination of
sections and categories for the long term model (α=1, β=1, χ=0, δ=0).
Table 1. Global averages (per user and per day) and relative increments in normalised precison
and recall for different configurations for the long term model for the selection process
The results shows (Table 1) that, in the case of the selection, the combination of
sections and categories with keywords is better, in normalised recall and precision,
than whatever other combination.
For the validation of the adaptation process, some experiment have been carried
out combining different possibilities for long term modeling, either acting on their
own or in combination with the short term model, S. For example, L(SC)S is the
combination of long and short term model when the long term model includes sec-
tions and categories (α=1, β=1, χ=0, δ=1).
Table 2. Global averages (per user and per day) and relative increments in normalised
precision and recall for different configurations for the long term model for the adaptation
process
It can be observed (Table 2) that the combination of long and short term is better
than if they are used alone, in terms of normalised recall and precision, and that the
best results in the combination of the short and long term models are obtained when
sections, categories and keywords are combined for the long term model.
References
1. Billsus, D. & Pazzani. M.J., 2000. “User Modeling for Adaptive News Access”, User
Modeling and User-Adapted Interaction Journal 10(2-3), pp. 147-180.
2. Díaz, A. & Gervás, P., 2004. “Adaptive User Modeling for Personalization of Web Con-
tents”. Proceedings of AH 2004, LNCS 3137, pp. 65-75.
3. Díaz, A. & Gervás, P., 2004. “Item Summarization in Personalization of News Delivery
Systems”. Text, Speech and Dialogue, TSD 2004, LNAI 3206, pp. 49-56.
4. Labrou, Y. & Finin, T., 2000. “Yahoo! As an Ontology: Using Yahoo! Categories to
Describe Documents”. Proceedings of CIKM-99, pp. 180-187, ACM Press.
5. Mizarro, S. & Tasso, C., 2002. “Ephemeral and Persistent Personalization in Adaptive
Information Access to Scholarly Publications on the Web”. 2nd International Conference
on Adaptive Hypermedia and Adaptive Web Based Systems, Málaga, España, 2002.
6. Salton, G., 1989. Automatic Text Processing: The Transformation, Analysis and Retrieval
of Information by Computer, Addison-Wesley Publishing, Reading, Massachussets, 1989.
7. Sebastiani, 2002. “Machine learning in automated text categorization”.
ACM Computing Surveys (CSUR), 34(1), pp. 1-47, ACM Press, 2002.
The gaps of the thesaurus WordNet used in
Information Retrieval
1 Introduction
Current Internet search engines try to solve the main problems that affect the quality of
the obtained results for getting a relevant collection of documents. Different users may
dis agree about the relevance of the obtained documents using the same query.
Relevance seems to be a subjective concept, mainly due to the ambiguity and the vo-
cabulary problems. Then, the search engine needs to know some kind of semantic
information that allows it to improve the results.
The main sources that are used to discover the semantic information are the dic-
tionaries, thesaurus or ontologies. WordNet [1] is one of the main tools used in infor-
mation retrieval processes, mainly for the meaning disambiguation tasks. This tool have
several problems such as, for example, the problem of the granularity of the senses or
the lack of recent terms. For example, WordNet 2.0 does not recognize the terms “XML”
or “CORBA”. Other times, it recognizes the common senses of one word, like in the
case of “SOAP”, but not recognizes a new sense (Simple Object Access Protocol).
This work is focused on the study of the weak points of WordNet. Our objective is
to use this information to improve our knowledge structures. For this reason, the study
of the user’s query is an important aspect that helps us to know the terms not recog-
nized by WordNet. This information can be later on used for getting better search
results.
2 GUMSe Overview
Meta-search engines offer a unified way to access multiple search engines, avoiding
the user to interact with all the underlying engines [2]. Taking advantage of the fea-
tures that they provide, we have developed GUMSe like a platform that allows us to
test new ideas applicable to the Web search process. The GUMSe architecture (Fig. 1)
is composed of a set of agents that communicate to each other by means of a network.
The main agents that form the architecture are Search Broker Agent, Sense Resolver
Agent, Query Agent Expansion, Rater Web Page Agent, Search Engine Agents, Con-
struction Ontology Agent and WordNet Agent.
3 Motivation
The aim of this work is to discover the user tendencies when they use the search en-
gines. With this information it is possible to design a more suitable evaluation algo-
rithm of the retrieved web pages. For this reason, we study the set of keywords and
queries more frequently used in the search engines.
There are many works focused on the user search techniques. Most of these works
show statistics about the number of keywords used, the number of queries per session,
and other statistical data [3]. Our objective is to discover the knowledge gaps of
WordNet. This information could be very useful to avoid negative situations in our
system. The main usual disadvantages are two: errors in the disambiguation process
and lack of information about the query terms.
Nowadays, the users often use the search engines to find the web page about a
concrete product, artist or program. In these cases, the traditional knowledge sources
fail because this information is about a specific domain, such as music or Internet. For
this reason, the construction of new knowledge structures that help the search engine
to know the specific domain of the user’s queries could be an important task. If the
search engine knows the domain, it is possible to modify some characteristics of the
search process. For example, if one user looks for one musical band and the system
knows it, the system could assign more importance to the results of one specialized in
music search engine. Another possibility is to introduce generic terms about the do-
main in the query expansion that avoid retrieved irrelevant web pages.
Usually, search engines offer tools for knowing the more consulted topics. Also,
there are many search engines and web pages that show the top searches. The data
used in our study are from a variety of sources. The main two are HitBrain1 and
Search.com2, among others.
4 Conclusions
In our preliminary study, WordNet recognizes the 45,4% of the more frequent used in
the user queries keywords. In this study, we used 10.000 keywords obtained from
HitBrain. The keywords have between 1 and 5 words. Another important issue is that
there are situations where WordNet recognizes keywords, but in a wrong way. These
cases are very infrequent but they produce mistakes in the search process.
It seems necessary to have good knowledge structures that can be used in informa-
tion retrieval to focus the search for obtaining good results. For this reason, the here
proposed subsystem, specialized in the improvement of the knowledge structures
could be a good support for Internet searchs.
Acknowledgements
Partially supported by PBC-03-004 PREDACOM project, JCCM, TIC2003-08807-C02-02
DIMOCLUST project, MCYT, and JCCM Cons. de Ciencia y Tecnología - European
Social Fund grants, Spain.
References
1. Miller, G.: Special issue, WordNet: An on-line lexical database, Intl. Journal of Lexicogra-
phy 3(4), 1990.
2. Gómez , M., Abasolo, J. M.: Improving meta-search by using query-weighting and nu-
merical aggregation operators. In Proc. of the Information Processing and Management of
Uncertainty conference, IPMU 2002, Annecy, France (2002): 1383-1390.
3. Jansen, B. J., Spink, A.: An Analysis of Web Documents Retrieved and Viewed. Interna-
tional Conference on Internet Computing (2003): 65-69.
1 http://www.hitbrain.com
2 http://www.search.com
Language for Uncertainty and Conflicting Agents
Extended Abstract
The traditional Object Oriented Technology is not sufficient agile at all to assume a
rapid development process. There is a lack of efficiency from the design to the
codification phase. This is due to the encapsulation of all the elements inside the
domain in one unique module: the class.
The RDM (Rapid Domain Modelling) architecture proposes another way for
transferring to code the different domain elements, following the Specialization
Principle and communication mechanisms through the Variable State Pattern. These
features are implemented in Entities, Relations and Operations, the main elements of
the RDM structure.
In this paper we want introduce a new dimensions inside the Object Oriented
language by the new language as the RDM. The new dimensions are uncertainty and
adaptation of the objects. Adaptation tries to reduce the uncertainty that came from
communication. Inside the OO we introduce new types of dynamic that balance the
uncertainty with the adaptation. Every agent perceives other agent’s behaviour as
coherent or incoherent in comparison with its particular behaviour (pattern). In the
communication among agents incoherence are possible. Incoherence means that for
the same object and the same property one agent give different and conflicting logic
valuation. A total coherence means that the property of the object is perceived from
the different agents from different points of view in the same way.
A set of True and False logic value for the same property give us the structure of
the uncertainty of the agent that can be formalise by special type of fuzzy logic
Adaptation or compensation try to change the structure of the uncertainty and also the
type of the fuzzy logic that agent use for the reasoning.
References
Extended Abstract
The design of knowledge-based systems has received special attention from the begin-
nings of Artificial Intelligence. In particular, the application of model-based systems has
obtained important results in the last decades [1]. However, the experience in develop-
ing these systems, far from becoming simpler, reveals the complexity on the design, the
development and its maintaining. Furthermore, in most of the cases, systems are highly
dependent on the domain and the problem to solve. The golden goal of a general model-
based architecture for all problem and domain is unreachable at the moment because of
the wide of the problem space to cover. However, general architectures fulfilling some
domain constraints could cover some parts of this problem space. Thus, in this work
we present a general model-based framework for intensive-knowledge domains, where
the temporal dimension (managing uncertainty) is considered to deal with the diagnosis
problem. The aim of this work is to present a general framework for temporal model-
based diagnosis. Nevertheless, the task of designing a general architecture is a complex
issue. One of the most difficult steps is to establish the set of goals to reach. Our general
framework tries to deal with the following key topics: domain independence, the man-
agement of temporal uncertainty, the knowledge acquisition bottleneck, model-based
representation, and the diagnosis task. In order to cover these points, firstly it is im-
portant to describe the main structure of the architecture, describing each part in detail
afterwards. This structure defines two modules, the Temporal/Domain Module and the
Decision Support Module.
The Decision Support Module includes all elements needed to obtain the diagno-
sis solution. This architecture is proposed for a model-based diagnosis. It is, therefore,
needed a formal model, which is used and stored in the knowledge base of the archi-
tecture. This knowledge base is supported by some knowledge acquisition tools, which
must access both the formal model and the domain ontology. Those tools are in charge
of the storage of the expert’s knowledge, and intended for the smooth of the problem of
the knowledge acquisition bottleneck. The diagnosis is the key process of the module.
This task, described in the architecture, is a temporal model-based diagnosis. To this
end, the role of the diagnosis agent is to find a diagnosis solution from: the temporal
data (from the temporal database), the domain ontology (from the ontology server), the
knowledge base, and the help of the temporal reasoner. This process is based on an al-
gorithm that obtains a diagnosis solution. This solution must be temporally consistent,
semantically consistent, fulfilling the constraints of the formal model, and explaining
the findings of the temporal database information.
Practical Experience.
This paper also relates our practical experience in the application of this architecture.
To this end, we choose the medical domain, in particular the Intensive Care Unit (ICU)
one.
ACUDES is a specific architecture that follows the proposed structure of this pa-
per (figure 1). Thus, it includes the Temporal/Domain Module. This module contains
the FuzzyTIME temporal reasoner and the Ontology Server, where the ICU ontology is
allocated. Furthermore, ACUDES also includes the Decision Support Module. There,
it is worth noting that we have designed and implemented CATEKAT, a web-based
tool to acquire the domain knowledge for the diagnosis model. Finally, the core of the
ACUDES diagnosis process is described by an abnormal behaviour model-based algo-
rithm. This model describes the causal and temporal behaviour of diseases. Thus, the
algorithm, from an abductive approach, obtains a diagnosis solution given the symp-
toms and the temporal behaviour model.
References
1. Console, L., Torraso, P.: A spectrum of logical definitions of model-based diagnosis. In
Hamscher, W., Console, L., de Kleer, J., eds.: Readings in Model-Based Diagnosis. Morgan
Kauffmann Publisher, Inc. (1992) 78–88
2. Campos, M., A.Cárceles, Palma, J., R.Marı́n: A general purporse fuzzy temporal information
management engine. In: Proceedings of the EurAsia-ICT 2002. (2002) 93–97
Ontology–Based Models of Human–Computer
Interaction for Usability Evaluation and Usable
Design
The work described here is founded on the analysis described by van Welie et al
[8], which establishes three knowledge domains: task domain, design domain and
user domain. In the basis of these domains and considering the usability defini-
tion proposed by ISO [3], we have structured the current version of our ontology
in six integrated (sub-)ontologies (see Figure 1). Aspects and usage indicators
of usability are described in the ’upper’ core concepts ontologies, and concepts
related with them are structured in two general partitions, one regarding the es-
sential interest point of usability, i.e. human-system interactions, and the other
regarding activities, methods and tools that are (left and right parts of Figure 1,
respectively). In consequence, concepts describing the physical fact of the inter-
action are separated from normative constructs about evaluating or designing-for
usability. The former include the application domain describing the structure of
the software-hardware application or system, the constructs describing human
users, their profiles and perceptual traits, and their purposeful interactions or
tasks, consisting ultimately on actions carried out by the users on the ’affor-
dances’ provided by the system. The latter includes norms and advices for the
design of user interfaces, and also methods, techniques and interpretations of
the way to evaluate and improve those designs, as a part of a usability-oriented
software development process or cycle. The application domain describes the
conceptual and physical structure of the user interface, and the human domain
describes the users and their abilities and cognitive processes. The task domain
models the dialogue and interaction between humans and systems, at different
levels of detail. The application domain is of special interest in some cases, e.g.
when considering mobile devices, due to the heterogeneity of their features that
require specific studies. The FIPA ontology2 provides a foundation for that kind
of specific domain. Design and evaluation are two interacting activities that serve
the purpose of making more ’usable’ interactions.
But it also allows for the specification of overall recommendations on the struc-
ture of applications through inference, e.g.:
LongHierarchyStructure ≡ HypermediaStructure u
∃linksT o.(HypermediaN ode u ∃linksT o.(HypermediaN ode u ∃linksT o.HypermediaN ode))
References
1. Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Patel-Schneider, P. (eds.).
(2003). The De-scription Logic Handbook. Theory, Implementation and Applica-
tions, Cambridge.
2. Berners-Lee, T., Hendler, J., Lassila, O. (2001). The Semantic Web. Scientific Amer-
ican, 284(5), pp. 34-43.
3. Bevan, N. (1998) ISO 9241: Ergonomic requirements for office work with visual
display terminals (VDTs) - Part 11: Guidance on usability. TC 159//SC 4 Technical
Committee of International Organization for Standardization.
4. Garcı́a, E., Sicilia, M. A., Gutiérrez, J.A. On the Vague Modelling of Web Page Char-
acteristics Regarding Usability. Atlantic Web Intelligence Conference (AWIC03),
Madrid, Spain. LNCS 2663. Springer, pp. 199-207. (2003).
5. Garcı́a, E., Sicilia, M.A., Hilera, J.R., González, L. Dialogue-Based Design Of Web
Usability Questionnaires Using Ontologies. Computer Aidded Design of User Inter-
faces IV, Kluwer Academic Press (to appear).
6. Scapin et al. (2000). A Framework for Organizing Web Usability Guidelines. 6th
Conference of Human Factors on the Web, Austin, TX.
7. Vanderdonckt, J., Development Milestones towards a Tool for Working with Guide-
lines, Interacting with Computers, Vol. 12, N2, 1999, pp. 81-118.
8. Van Welie, M., van der Veer, G.C. & Eliëns, A. (1999), Breaking down usability,
Proc. of Interact’99. 613-620.
4
http://usability.gov/guidelines/
Extension of Ontologies assisted by Automated
Reasoning Systems?
Extended Abstract
References
1. J. A. Alonso-Jiménez, J. Borrego-Dı́az, A. M. Chávez-González and J. D. Navarro
Marı́n, Towards a Practical Argumentative Reasoning with Qualitative Spatial
Databases, 6th International Conference on Industrial and Engineering Applica-
tions of Artificial Intelligence and Expert Systems, Lecture Notes in Artificial In-
telligence vol. 2718, 789-798, Springer-Verlag Berlin, 2003.
2. J. A. Alonso-Jiménez, J. Borrego-Dı́az and A. M. Chávez-González, Ontology
Cleaning by Mereotopological Reasoning. Proceedings of DEXA Workshop on Web
Semantics (WebS’04). IEEE Press, pp. 137-137, 2004.
3. J. A. Alonso-Jiménez, J. Borrego-Dı́az and A. M. Chávez-González, Foundational
challenges in Automated and Ontology Cleaning in the Semantic Web. To appear
in IEEE Intelligent Systems.
4. A. G. Cohn, B. Bennett, J. M. Gooday and N. M. Gotts. Representing and Rea-
soning with Qualitative Spatial Relations about Regions. Chapter 4 in O. Stock
(ed.), Spatial and Temporal Reasoning, Kluwer, Dordrecth, 1997.
5. B. Bennett, The Role of Definitions in Construction and Analysis of Formal On-
tologies, in: P. Doherty, J. McCarthy, and M. Williams (eds.) Logical Formalization
of Commonsense Reasoning (2003 AAAI Spring Symposium), 27-35, AAAI Press,
2003.
6. A.M. Chávez-González, Mereotopological Automated Reasoning for Ontology
Cleaning, forthcoming Ph.D. thesis.
Applying Genetic Algorithms to Referring Expression
Generation and Aggregation
1 Introduction
Natural Language Generation (NLG) includes complex tasks [2], and features such as
ambiguity can make these tasks too complex in fact to be tackled with classic search
algorithms in a reasonable time. Evolutionary Algorithms (EAs) are heuristic tech-
niques [1] based on a search model which emulates evolutionary phenomena, and that
can be used to speed up the search. The purpose of most EAs is to find a good solu-
tion but not necessarily the best solution, and this is enough for most natural lan-
guages processes. In this paper we present a first approach to the idea of using NLG
and EAs together.
The specific problems addressed in this paper are two subtasks of the process of
automatically generating a natural language text for a given content: referringf expre-
sion generation and aggregation. The nature of these subtasks is best illustrated over
an example:
A princess lived in a castle. She loved a brave knight. She was
pretty. The castle had towers. It was strong. They lived in it.
In this text various elements of the reality that is being described are mentioned
one or more times: a princess, a knight,... Different natural language expressions are
used to describe them at each occurrence in the text: “a princess”, “she”, “the prin-
cess”. The task of referring expression generation concers the way in which a system
decides at each occurrence how to refer to a given element.
The correct use of referring expressions to compete with human generated texts
involves a certain difficulty [4]. Possible simple algorithms for deciding when to use
a pronoun and when to use the full name produce poor results. Two occurrences of
the same concept in a paragraph can be far apart, and this may confuse the reader.
Knowledge intensive approaches modelled on the way humans do it require a certain
measure of content understanding that is resource hungry.
Aggregation [3] involves deciding how compact the presentation of information
should be in a given text. It operates at several linguistic levels, but we only consider
it here with respect to concepts and their attributes. For instance, the system must
decide between generating “The princess is blonde. She sleeps.” and generating “The
blonde princess sleeps.”. Aggregation is generally desirable, but may result in adjec-
tive-heavy texts when the information to impart becomes dense in terms of attributes,
as in “The pretty blonde princess lived in a strong fancy castle with her stern rich
parents.”. A good solution should find a balance between conciseness and acceptable
style.
2 An Evolutionary Solution
We propose the use of a simple genetic algorithm, where the appearances of the con-
cepts are considered as the genes. The initial population is generated randomly, using
for each concept its full name or its pronoun. When using the full name, a selection of
the attributes the concept has in the knowledge base is chosen. These attributes will
appear just before the name of the concept, as it is usual in English. The system
works over this population for a number of generations determined by the user. In
each generation three genetic operators are used: crossover, mutation and aggrega-
tion. Finally, at the end of each generation each tale is evaluated and a selection of the
population is passed to the next one, in such way that the tales with a higher fitness
have more possibilities of being chosen.
The operators
For the crossover operator, two drafts are selected randomly and crossed by a ran-
dom point of their structure. So, each of the sons will have part of each of the parents.
In the case of the mutation operator, some of the genes are chosen randomly to be
mutated. If the gene is a pronoun, it will change into the correspondent full name,
always associated with a subset of its possible attributes. In case the gene was a full
name, there are two options: to change it into a pronoun, or to change the subset of
attributes that appear with it. One of these two options is chosen randomly.
The aggregation operator addresses the task of deciding on the aggregation be-
tween concepts and their attributes. This involves a certain modification of the struc-
ture of the text, because sentences in the text may be deleted if the information they
impart becomes part of a previous sentence. The aggregation operator acts only on
genes corresponding to explicitly mentioned concepts: concepts referred by pronouns
are excluded. It can act in two directions. If the reference to the concept appears with
one or more attributes – as in “A blonde princess lived in a castle.” -, the operator
disggregates the attributes by eliminating their mention and adding a corresponding
“X is Y” sentence – resulting in “A princess lived in a castle. She was blonde.”. If the
reference to X has no attributes – as in “A princess lived in a castle.” -, the algorithm
looks for an “X is Y” sentence – such as “The princess was blonde.”-, adds the cor-
responding attributes to the reference, and deletes the “X is Y” sentence – resulting in
“A blonde princess lived in a castle.”.
The goal of this definition of the aggregation is to ensure that the attributes of a
concept be mentioned in the appearance of a concept or in the correspondent “X is Y”
sentences, but not in both. As the aggregation operator are used randomly, the desired
result is obtained only in some cases.
The fitness function
The key to the genetic algorithm lies in the choice of fitness function. A simple ap-
proach would be to require that in each generation the user reads all the texts and
gives them a fitness value. The number of generations and individuals in the popula-
tion for a simple experiment makes this approach impractical. So, we analyzed the
features of a human generated text from the point of view of the referring expres-
sions, and we found five different evaluation functions, each one aimed at a different
aspect.
Correct Referent. When writing a text, we cannot use a pronoun for something
that we have not mentioned before, or readers would get confused. In addition, if the
full name reference and the pronoun are far, the reader can also get confused and be
unable to link the two occurrences of the same concept. So, this function penalizes
the appearances of pronouns for a concept that has not been referred in the two previ-
ous genes with its full name.
Redundant Attributes. When describing a concept in an “X is Y” sentence, peo-
ple do not use the attribute they are going to describe in the reference to the concept.
This function penalizes sentences like “The blonde princess was blonde”.
Reference Repetition. Using always the same reference together with the same at-
tributes results in repetitive text. For example, it is acceptable to use “the princess”
every time we refer to the princess character, but it would be striking to use always
“the pretty blonde princess”. To avoid that, repetitive use of references is penalized.
Coherence. If we use different subsets of its attributes in different references to
the same concept, the reader may mistakenly assume that we are referring to different
concepts. For example, if we use “the pretty princess” and “the blonde princess” in
different places, and we have not specified that the princess is both pretty and blonde,
it could seem that there are two princess: a pretty one and a blonde one. With this
function, we penalized the appearance of attributes that have not been mentioned in
the first description of a concept or in a “X is Y” sentence.
Lack of Concept Information. When applying the genetic operators to a draft,
some information about a concept may disappear from the final text. To avoid that,
this function looks for all the attributes of each concept in the ontology, and checks if
they are mentioned in the text. If not, the draft is penalized.
For the evaluation of each of the texts that form the population, we use formula
(1), where error is the accumulated error of a specific text taking into account the
fitness functions mentioned above.
fitness = 1 / ( error + k ). (1)
In this way, the fitness would be greater if the error is smaller. The constant k is
used to avoid divisions by zero. In our experiments it was set with the value 1, so the
maximum possible fitness was 1.
The results
The following text is an example of system output for a very simple input concerning
the initial description of the characters that take part in a tale:
A princess lived in a castle. She was the daugther of parents. They
were stern. The blonde princess loved a brave handsome knight. The cas-
tle had towers. The parents lived in the castle. It was strong.
The beginnings of stories are rich in descriptions of characters and locations and
are therefore good case studies for the problems being discussed. The input to the
system provides conceptual descriptions of characters and locations in terms of at-
tributes and relations between them. The evolutionary algorithm operates over the
output of previous stages that have carried out other NLG tasks such as content de-
termination – selecting the particular concepts that are relevant – and discourse plan-
ning - organising them in an orderly fashion. These tasks are currently carried out in a
traditional manner and simply provide the data for the evolutionary stages. A tradi-
tional surface realization stage transforms the output of the evolutionary stages into
readable text.
With respect to both of the tasks addressed, the output texts respect the strong con-
straints required for the text to be acceptable, while at the same time showing reason-
able variation between the different options much as a human-generated text would.
Due to the main difficulties in the evaluation of the population at the end of each
generation, we are currently considering the possibility of using neural networks to
reproduce the human evaluation of the quality of the resulting texts. We are also
working on extending the system to allow the use of proper nouns to describe some
concepts, as an additional option to pronouns and descriptive references, including
the revision of the genetic operators and the introduction of new evaluation functions
to estimate the correct application of proper nouns. In view of these results, in future
work we want to apply EA techniques to other tasks of NLG, such as content deter-
mination and discourse planning. The particular advantages of evolutionary tech-
niques, combined stage by stage in this manner, may be an extremely powerful
method for solving natural language generation problems while also profiting from
classic NLG techniques.
References
1. Holland, J. H., Adaptation in Natural and Artificial Systems. An Introductory Analysis with
Applications to Biology, Control and Artificial Intelligence, MIT Press, Cambridge, Massa-
chusetts, Second Edition, 1992.
2. Reiter, E. “Has a consensus NL generation architecture appeared, and is it psychologically
plausible?”, in: McDonald, D. and Meteer, M. (eds.) Proccedings of INLGW ’94, pages
163-170, Kennebunkport, Maine, 1994.
3. Reape, M, Mellish, C., “Just what is aggregation anyway?”, in: Proceedings of the 7th Euro-
pean Workshop on Natural Language Generation, Toulouse, France, 1999.
4. Reiter, E. Dale, R. “A fast algorithm for the generation of referring expressions”, in: Pro-
ceedings of the 14th conference on Computational linguistics - Volume 1, Nantes, France,
1992
!!! "#$%
"& $''$
$ (
)*+ (,!%" ,
!
" #$
# %
-
..#
" /0 (
& 1
2
$''$
# 3%4 (%5 6 6%6 %7
8
%
8
'
".)$''"*+9%" 6
$''"
& 5
:
, 7
8
%
( &
&'
)
#
# $
#$ "'&%"$ $''#
; < 3 /0 ( ' )*
"
$ /
1
7
5
...
An Ontology for Reusing Synthetic Tasks
Extended Abstract
Table 1. We use the task types as they are defined in the CommonKADS’s book [8]
CommonKADS and authors like [6] suggest a shared characterisation of all syn-
thetic tasks by using some terms such as requirements, constraints, domain, etc. These
terms are similar to those used by the Constraint Satisfaction (CS) techniques [2],
such as the Forward Checking algorithm. We have adopted this technique (Figure 1
shows de derived PSM), because it provides us with the needed simplicity to trans-
form it into a PSM and is easily extensible with some other techniques such as consis-
tency techniques and variable/value ordering. A detailed description of this PSM can
be found in [7].
The full paper will describe the roles of the PSM (i.e. task description knowledge,
soft requirement, etc.) and the association between the domain vocabulary and some
of the test applications.
According to the people participating in this initiative, the reuse of the PSM and its
associated executable code supposed at least 30% of development effort saving (just
for the expert module). This percentage is even larger depending on some other fac-
tors such as the complexity of the application being modelled. The PSM Ontology
allow us to prepare structured interviews with the human experts in order to acquire
specific knowledge for that role. The experts easily understood the PSM diagram and
collaborated actively providing valid rules that were later included in the system.
Task
Input Data Generate Composition
Knowledge
Task
Opera- Partial
Description
tionalize Solution
Knowledge
Soft
Requirement
Propose Undo
Extension Extension
Preferences
Ordering
Knowledge
Extension Violation
On the other hand, assessing a priori the potential benefits of reusing components
is not easy because there are too many (obscure) factors that directly influence on the
final revenue. Some of these include the technological background and experience of
the development team, the size and complexity of the task type and domain applica-
tion, and how easy the adaptation of the module will result to be reused to the actual
domain problem.
References
1 Angele, J., Fensel, D., Studer, R. Domain and Task Modelling in MIKE. In Suthliffe, A et al.
(eds). Domain Knowledge for interactive system design. Chapman & Hall. 1996
2 Baptiste, P., Le Pape, C., and Nuijten, W., Constraint-Based Scheduling. Applying con-
straint programming to scheduling problems. Kluwer Academic Publishers. 2001
3 Chandrasekaran, B., Josephson, J., Benjamins R. Ontologies: What are they? Why do we
need them? IEEE Intelligent Systems and Their Applications, 14(1): 20-26, 1999. Special
issue on Ontologies.
4 Domingue, J., Motta, E., Watt, S. The emerging VITAL workbench. In Aussenac et al.
editor, EKAW’93 Knowledge Acquisition for Knowledge-Based Systems. Lecture Notes in
Artificial Intelligence, LNCE723, Berlin Germany, 1993, Springer-Verlag.
5 Gómez-Pérez, A., Benjamins, R. Overview of knowledge sharing and reuse components:
Ontologies and Problem-Solving Methods. Proceedings of the IJCAI-99 workshop on
Ontologies and Problem-Solving Methods (KRR5) Stockholm, Sweden, 1999
6 Motta, E., Rajpathak, D., Zdrahal Z., Roy, R. The Epistemology of Scheduling Problems.
Proceedings of the 15th European Conference on Artificial Intelligence, F. van Harmelen
(editor), IOS Press, Amsterdam, 2002
7 Rodríguez, A., Palma, J., Quintana, F. Experiences in Reusing Problem Solving Methods.
An application to Constraint Programming. Knowledge-based intelligent information and
engineering systems. LNAI 2774, Springer, 2003.
8 Schreiber, Akkermans, Anjewierden, de Hoog, Shadbolt, Van de Velde, Wielinga. Knowl-
edge Engineering and Management. The CommonKADS Methodology. The MIT Press.
1999
Modelling and Simulation of Systems with
Multi-level Anticipation and Retardation
The state of the first operational level is given by the temporal function P(t),
which may represent the production of the company. The state of the second
strategic level is given by F(t) representing, for example, Research and Devel-
opment. The state of the third level, at last, is given by V(t), representing the
normative work of defining the company’s vision and mission. The three func-
tions Gi are functions of the three states P (t) , F (t) and V (t) and the sets of
parameters given by pi .
The model as it is, is a model describing the company at the current time
t. The numerical simulation of this model can be performed by giving explicitly
the functions Gi , the initial states at time t = 0 for P (0) , F (0) and V (0) and
the values of the parameters pi . Let us now give an example of retardation and
one of anticipation:
References
1. Dubois Daniel M.:Mathematical foundations of discrete and functional systems with
strong and weak anticipations. In Anticipatory Behavior in Adaptive Learning Sys-
tems, State-of-the-Art Survey, Edited by Martin Butz et al, Lecture Notes in Ar-
tificial Intelligence, Springer LNAI 2684. (2003) 110–132
2. Espejo Raul, Schuhmann Werner, Schwaninger Markus, Bilello Ubaldo: Organi-
zational Transformation and Learning, A Cybernetic Approach to Management.
Wiley, Chichester. (1996) 227–270
3. Holmberg Stig C.:Multimodal Systemic Anticipation. I. J. of Computing Anticipa-
tory Systems, Vol. 3. (1999) 75–84
A Tractable Subclass of Fuzzy Constraint Networks
1 Introduction
A Fuzzy Constraint Network (FCN) Ld consists of a set of n+1 variables X0, ... ,Xn
whose domain is a full ordered set of equidistant precise values, and a finite set of
disjunctive binary constraints among these variables. The variables and constraints
can be represented with a constraint graph or network. The complete model definition
is included on [2].
A network is series-parallel if it can be reduced to an arc applying iteratively a
reduction operation. The SP algorithm [2] decides whether a network is series-
parallel, using only the graph topology, but not the information contained into
network constraints. Enhancing this algorithm to use constraint information, we can
obtain useful information about the network, and to decide simultaneously whether
the network is series-parallel or not.
Working in this direction, we propose the algorithm SP+, which can decide with an
O(n) complexity whether a network is inconsistent if it detects an empty constraint. If
an empty constraint is not detected and the algorithm decides that the network is
series-parallel, the network is consistent. As a subproduct, the final obtained
constraint is the minimal constraint among the two lasting variables, and the
equivalent network obtained with the computed constraints is decomposable in the
reverse variable reduction ordering. This is the code of SP+ algorithm:
SP+ Algorithm
begin
for each i=0..n Compute-degree (i)
NodeQueue = {nodes with degree 1 and 2}
while (NodeQueue <> ∅ and |V| > 2)
begin
node=Get(NodeQueue)
if Degree(node) = 2 then Lij ← Lij ∩
&Lik ⊕
&L
kj
if Lij = ∅ then exit
V ← V − {node}
if Degree(node)=1 then Degree(Neighbour(node)) --
if Degree(Neighbour(node)) = 2 then
Put(NodeQueue,Neighbour(node))
else if Connected(Neighbours(node)) then
Degree(Neighbour(node)) --
if Degree(Neighbours(node)) = 2 then
Put(NodeQueue, Neighbours(node))
else E ← E + {ArcNeighbours(node)}
end
if (NodeQueue = ∅ and |V| > 2) then
exit (“Network is not series-parallel”)
end
Fig. 1. Example of network to be reduced with SP+
Figure 1 shows a FCN network to be reduced with SP+. Figure 2 shows the
reduction process and refined constraints obtained.
References
1. A. Bosch, M. Torres, and R. Marín. Reasoning with Disjunctive Fuzzy Temporal Constraint
Networks. In Proc. of 9th Int. Symposium on Temporal Representation and Reasoning
(TIME’02). IEEE Computer Society Press, pages 36-43, 2002.
2. A. Bosch, F. Guil, C. Martinez, and R. Marín. Series-Parallel and Tree-Decomposition
Approaches for Fuzzy Constraint Networks. In Proc. of the 8th Ibero-American Conference
on Artificial Intelligence (IBERAMIA’02). LNAI, 2527:275-284, 2002.
3. R. Dechter, I. Meiri, and J. Pearl. Temporal constraint networks. Artificial Intelligence,
49:61-95, 1991.
4. R. Marin, S. Barro, A. Bosch, and J. Mira. Modeling time representation from a fuzzy
perspective. Cybernetic & Systems, 25(2):207-215, 1994.
Parallel State Space Generation and Exploration
on Shared-Memory Architectures
Milan Češka, Bohuslav Křena, Tomáš Vojnar
1 Introduction
Model checking is a technique for an automated formal verification of correctness of
various classes of systems. Compared to the more traditional approaches to this prob-
lem based on simulation and testing, the main advantage of model checking is that the
desirable behavioural properties of the considered system are checked in such a way
that all the reachable states which may affect the given property are guaranteed to be
covered. The disadvantage of model checking is that it may have very high compu-
tational time and space requirements. That is why various ways of dealing with the
computational complexity of model checking are sought.
The systems to be verified may be described using various languages. Here, we
consider the systems to be described by the PNtalk language based on Object-Oriented
Petri Nets (OOPNs). PNtalk [6] has been developed at the Brno University of Technol-
ogy as a tool suitable for modelling, prototyping and rapid development of concurrent
and distributed software systems.
In the paper, we discuss possibilities of parallel OOPN state space generation and
exploration on shared-memory architectures. The goal is to combat the high time com-
plexity of state spaces-based verification methods.
NT NT
t>0 t=0 transition guard
push NT:=t+1 NT:=t-1
(NT,x) (t,x)
#ok (#ok,x) #err
PSfrag replacements
method net
output stack
return return
place
Application is_a PN
y
x (#ok,x)
p5 p4
t4
Node
Thread
hash table
marking pool
PSfrag replacements
As we have already said, one of the crucial issues for efficiency of parallel applica-
tions is the quality of load sharing among threads. We have started with one data node
(i.e. a fraction of the hash table) for each thread. However, this did not work well. At
first, we are not able to divide the hash table to parts with the same number of states
in advance. Moreover, for a fine load sharing, it is necessary that each thread has some
work (i.e. states to explore) during the whole production run, which is impossible to
predict and guarantee in advance too.
Therefore, we have divided the hash table and the work list to more parts than we
have threads. We find by experiments that the best performance is reached when three
times more nodes than threads are created. Then, threads switch among nodes and try
to find some work to do as it is shown in Fig. 3 for the case of two threads.
Using all these optimizations, we have reached parallel speedups up to 3.3 using 16
processors on the Sun Fire 15k server and speedups up to 2.3 using 4 processors on the
Sun Enterprise 450 server. Moreover, the work distribution among particular threads is
fine (e.g., for 4 threads and about 10 million states: 24.3 % – 25.5 % – 25.3 % – 24.3 %).
The main problem that prevents us to achieve an even better parallel speedup is
the memory management system (especially garbage collecting) due to the memory
operations being performed sequentially. A possible approach to this problem is to im-
plement a user-specific memory management system like in [1], to reduce the number
of dynamically created objects.
N0 N1
Thread 0
N5 N2
PSfrag replacements
Thread 1
N4 N3
In the full paper, we describe the proposed algorithm in more detail including a dis-
cussion of efficient ways of storing and comparing the highly complex states of OOPNs,
more experimental results, etc.
Acknowledgement. This work was supported by the Czech Grant Agency under the
contracts 102/04/0780 and 102/03/D211 and it was also supported by the European
Community within the “Access to Research Infrastructure Action of the Improving Hu-
man Potential Programme” under the contract HPRI-CT-1999-00026.
References
1. S. C. Allmaier and G. Horton. Parallel Shared-Memory State-Space Exploration in Stochas-
tic Modelling. In Proc. of IRREGULAR ’97, LNCS 1253, 1997. Springer.
2. M. Bull and M. E. Kambites. JOMP—an OpenMP-like Interface for Java. In Proc. of the
ACM 2000 conference on Java Grande, 2000. ACM Press.
3. E. M. Clarke, O. Grumberg, and D. A. Peled. Model Checking. The MIT Press, Cambridge,
Massachusetts, London, England, 2000.
4. M. Češka, V. Janoušek, and T. Vojnar. Towards Verifying Distributed Systems Using Object-
Oriented Petri Nets. In Proc. of EUROCAST’99, LNCS 1798, 2000. Springer.
5. M. Češka, L. Haša, and T. Vojnar. Partial Order Reduction in Model Checking of Object-
Oriented Petri Nets. In Proc. of EUROCAST’03, LNCS 2809, 2003. Springer.
6. V. Janoušek. Modelling Objects by Petri Nets. PhD. thesis, Brno University of Technology,
Brno, Czech Republic, 1998.
7. B. Kočı́. Methods and Tools for Implementing Open Simulation Systems. PhD. thesis, Brno
University of Technology, Brno, Czech Republic, 2004.
8. B. Křena. Analysis Methods of Object Oriented Petri Nets. PhD. thesis, Brno University of
Technology, Brno, Czech Republic, 2004.
9. T. Vojnar. Towards Formal Analysis and Verification over State Spaces of Object-Oriented
Petri Nets. PhD. thesis, Brno University of Technology, Brno, Czech Republic, 2001.
Tree-structured Legendre multi-wavelets
1 Introduction
Recently, multi-wavelets, that is, wavelets with more than one scaling and wavelet
function have gained a considerable interest in the area of signal processing, in
particular in signal de-noising and compression due to their promising results in
these areas.
The first multi-wavelets were introduced by Alpert [1], who constructed
Legendre type of wavelets on the interval [0,1) with several scaling functions
φ0 , φ1 , . . . , φN −1 , instead of just one scaling function φ0 . This difference enabled
high-order approximation with basis functions supported on non-overlapping in-
tervals. In the particular case of N = 1, Alpert’s multi-wavelets coincide with
the Haar basis.
The discrete analogue of continuous Legendre multi-wavelets has been also
introduced by Alpert [1]. The structure of this analogue is essentially similar
to that of continuous multi-wavelet bases, but the discrete construction is more
convenient when the representation of a function (and its related operations) is
based on its values at a finite set of points [1].
In [3], a new basis, called the tree-structured Haar (TSH) basis was intro-
duced, which is a generalization of the classical Haar basis to the arbitrary time
and scale splitting. The idea behind such a construction was to adapt the basis
to the signal on hand. Discrete orthogonal TSH transform has been successfully
applied to the problem of de-noising signals [5].
In [4], the TSH structure has been extended to the multi-wavelet bases by
the analogy with Alpert’s construction. The construction of the basis functions
assured their orthogonality in the continuous space. However, the question re-
garding the construction of discrete orthogonal bases was left unexplored. In this
paper we address the problem of constructing the discrete counterpart of tree-
structured multi-wavelets, so called, tree-structured Legendre (TSL) transform,
by generalizing the construction procedure, introduced in [1].
3 Discrete counterparts
2.5 3
2.5
2
2
1.5
1.5
1
1
0.5 0.5
0
0
−0.5
−0.5
−1
−1
−1.5
−1.5 −2
0 0.5 1 0 0.5 1
(a) (b)
⎜ 1 xui +2 . . . x2N −1 ⎟
⎜ ui +2 ⎟
M1,i =⎜ . . ⎟, (2)
⎝ .. .. ⎠
2N −1
1 xui +2N . . . xui +2N
if c(d − 1, i) is a splitting node, and
⎛ ⎞
1 xui +1 . . . x2N
ui +1
−1
⎜ 1 xui +2 . . . xu +2
2N −1 ⎟
⎜ i ⎟
M1,i = ⎜ . .. ⎟, (3)
⎝ .. . ⎠
1 xui +N . . . x2N −1
ui +N
i−1
if c(d − 1, i) is a non-splitting node. Here ui = j=1 ν(c(d − 1, j)) for i >
1, and u1 = 0. Additionally, we define matrices U1,i = (Orth(M1,i ))T , i ∈
{1, 2, . . . , µd−1 }, where Orth stands for column-by-column Gram-Shmidt orthog-
onalization.
By S U and S L let us denote the upper and lower k rows of (2k×2k) matrix S,
respectively. The first (n × n) basis matrix, U1 , is obtained by first constructing
an auxiliary matrix, Ũ1 :
⎛ L ⎞
Ũ1,1
⎜ L
Ũ1,2 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ L ⎟
⎜ Ũ1,µd−1 ⎟
U˜1 = ⎜ U ⎟, (4)
⎜ Ũ1,1 ⎟
⎜ ⎟
⎜ U
Ũ1,2 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
U
Ũ1,µ d−1
where ⎧
, if c(d − 1, i) is a splitting node
⎨ U1,i
Ũ1,i = IN (5)
⎩ , otherwise.
0N
12
8 4
4 4 4
4 2 2 4
2 2 2 2 2 2
Ũj,i and Uj,i are related in the same way as Ũ1,i and U1,i (see (5)), Uj,i =
(Orth(Mj,i ))T , and Mj,i is given by
U
Uj−1,k × Mj−1,k
Mj,i = (8)
U
Uj−1,k+1 × Mj−1,k+1
if c(d − j, i) is a splitting node and c(d − j + 1, k), c(d − j + 1, k + 1) are its
left and right children, respectively; and Mj,i = Mj−1,k if the node c(d − j, i) is
non-splitting, and c(d − j + 1, k) is its only child.
The final matrix, U = Ud × Ud−1 × · · · × U1 , represents the discrete wavelet-
like basis of parameter N on x1 , x2 . . . , xn . As it has been mentioned in [1], some
adjustments must be made to this formula to ensure numerical stability. These
issues are discussed in [2] and we will not cover them here.
References
1. Alpert, B.: A class of bases in L2 for the sparse representation of integral operators,
SIAM J. Math. Anal., 24, (1993), 246–262
2. Alpert, B., Beylkin, G., Coifman, R., Rochlin, V.: Wavelet bases for the fast solution
of second-kind integral equations, SIAM Journal on Scientific Computing, 14, 1,
1993, 159–184
3. Egiazarian K., Astola, J.: Tree-structured Haar transforms, Journal of Mathematical
and Imaging Vision, 16, 2002, 267–277
4. Gotchev, A., Egiazarian, K., Astola, J.: On tree-structured legendre multi-wavelet
bases, Proc. Int. Workshop on Spectral Methods and Multirate Signal Processing
(SMMSP 2001), June 16-18, Pula, Croatia, 2001, 51–56
5. Pogossova , E., Egiazarian, K., Astola, J.: Signal de-noising in tree-structured Haar
basis, Proc. 3rd International Symposium on Image and Signal Processing and Anal-
ysis (ISPA’03), Rome, Italy, September 18-20, 2, 2003, 736–739
Remarks on Calculation of Autocorrelation on
Finite Dyadic Groups by Local Transformations
of Decision Diagrams
1 Introduction
Autocorrelation is an important operation in signal processing and systems the-
ory [1]. In particular, the autocorrelation on finite dyadic groups, denoted as
dyadic autocorrelation Bf , (see Definition 1) is useful in switching theory and
design of systems whose inputs and outputs are represented by functions defined
in 2n , n ∈ N points, including switching functions as an example [2], [3], [6],
[7], [8], [9]. Recently, some new applications of dyadic autocorrelation in spectral
methods for switching functions, testing of logic networks [4] and optimization
of decision diagrams (DDs) for representation of discrete functions has been re-
ported [10]. In this paper, we define and discuss a method for calculation of
the dyadic autocorrelation through decision diagrams, the use of which permits
processing of functions of a large number of variables.
2 Background Theory
Denote by C2n the finite dyadic group, where C2n = ×ni=1 C2 , and C2 = ({0, 1}, ⊕),
where ⊕ denotes multiplication modulo 2 (EXOR).
Definition 1 For a function f : C2n → R, where R is the field of real numbers,
P2n −1
the autocorrelation Bf is defined by Bf (τ ) = x=0 f (x)f (x ⊕ τ ), where τ =
0, . . . , 2n − 1. In binary notation, x = (x1 , . . . , xn ) and τ = (τ1 , . . . , τn ), where
xi , τi ∈ {0, 1}.
In matrix notation, if f and Bf are represented by vectors F = [f (0), . . . , f (2n −
1)]T and Bf = [Bf (0), . . . Bf (2n − 1)]T , respectively, then, Bf = B(n)F, where
B(n) is the dyadic autocorrelation matrix for f .
Bf = 2n W −1 (W f )2 ,
3 Decision Diagrams
Decision diagrams are data structures providing compact representations of dis-
crete functions defined in a large number of points [11]. In this paper, we as-
sume that a given function f is represented by a Multi-terminal Binary DD
(MTBDD(f )) [11]. A MTBDD is a directed acyclic graph consisting of non-
terminal nodes and constant nodes connected by edges. Each node has two out-
going edges labeled by the negative and positive literals xi and xi of the decision
variable assigned to the node. Nodes to which the same variable is assigned form
a level in the MTBDD.
Notice that in calculations over decision diagrams, the impact of the deleted
nodes should be taken into account through the cross points defined as points of
intersections of paths from the root node to the constant nodes with the imag-
inary lines showing levels in decision diagrams, which means lines connecting
nodes to which the same decision variable is assigned [12]. Complexity of a de-
cision diagram is usually expressed in terms of the number of non-terminal and
constant nodes, called the size of the decision diagram.
f f Bf
_ S2 _ S2 _ 2S
x1 x1 x1 x x1 x1
1
S2 S
_ S2 _
S2
_ 2
x2 x2 x2 x2 _ x2 x2 x2
x2
_ S2 _ S2 _ S2 _ S2 S2 _ S2
x3 x3 x3 x3 x3 x3 x3 x3 x3 x3
0 0 1 2 3 3 3 3 0 1 2 3 41 40 36 18
Example 1 Fig. 1 shows a MTBDT and the corresponding MTBDD for a func-
tion f given by the vector of function values F = [0, 0, 1, 2, 3, 3, 3, 3]T . In this
figure, we also show the cross points in the MTBDD. Fig. 1 shows the MTBDD
for the autocorrelation function Bf (τ ).
References
1. Agaian, S., Astola, J., Egiazarian, K., Binary Polynomial Transforms and Nonlin-
ear Digital Filters, Marcel Dekker (1995).
2. Karpovsky, M.G., Finite Orthogonal Series in the Design of Digital Devices, John
Wiley (1976).
3. Karpovsky, M.G., (ed.), Spectral Techniques and Fault Detection, Academic Press
(1985) 35-90.
4. Karpovsky, M.G., Stanković, R.S., Astola, J.T., Spectral techniques for design and
testing of computer hardware”, Proc. Int. Worksop on Spectral Techniques in Logic
Design, SPECLOG-2000, Tampere, Finland, June 2-3 (2000) 1-34.
5. Meinel, Ch., Somenzi, F., Tehobald, T., Linear shifting of decision diagrams and
its application in synthesis, IEEE Trans. CAD, Vol. 19, No. 5 (2000) 521-533.
6. Pichler, F., Walsh functions and linear system theory, Proc. Applic. Walsh Func-
tions, Washington, D.C. (1970) 175-182.
7. Pichler, F., Realizations of Prigogine’s Λ-transform by dyadic convolution, in
Trappl, R., Horn, W., (eds.), Austrian Society for Cybernetic Studies, ISBN
385206127X (1992).
8. Rice, J.E., Muzio, J.C., Methods for calculating autocorrelation coefficients, Proc.
4th Int. Workshop on Boolean Problems, (IWSBP2000) (2000) 6976.
9. Rice, J.E., Muzio, J.C., Use of autocorrelation function in the classification of
switching functions, Euromicro Symposium on Digital System Design (2002) 244-
251.
10. Rice, J., Serra, M., Muzio, J.C., The use of autocorrelation coefficients for variable
ordering for ROBDDs, Proc. 4th Int. Workshop on Applications of Reed-Muller
Expansion in Circuit Design, Victoria, Canada, August 20-21 (1999) 185-196.
11. Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Acad-
emic Publishers (1996).
12. Stanković, R.S., Sasao, T., Moraga, C., Spectral transform decision diagrams, in T.
Sasao, M. Fujita, (ed.), Representations of Discrete Functions, Kluwer Academic
Publishers (1996) 55-92.
A new Pseudo-Random Generator based on Gollmann
Cascades of Baker-Register-Machines
Extended Abstract
c
111…11
….
…. …. ….
An analysis of the CBRM shows that it has good statistical properties. For example
n-gram distributions are uniform.Another nice property of a BRM is that the
generated sequences are highly sensitive to the initial key. By statistical analysis, we
found that changing one bit in the encryption key caused 49.4% of the bits to change
in the corresponding cipher-image. Optimally, fully half of the bits should change, in
order to hide any information about the key from leaking.The linear complexity
profile of the CBRM algorithm is staying very close to the length of the generated
sequence.The cycle length (period) of the CBRM depends on the partition key, as
mentioned above. In general it is difficult to compute the exact cycle length of the
CBRM. A general formula for this computation can be given as follows:
Let r be the number of the feedback-shift-registers that realize the Baker-Register-
Machine, σ i (i =1, 2, …, r) the length of the shift-register i, and k the number of
cascaded stages. Then the cycle length of the CBRM is given by:
L = lcm (σ 1 , σ 2 ,...., σ r )k
References
Josef Scharinger
1 Introduction
Permutations are a core component of almost every cipher. No matter if we
consider the DES [5], AES [2] or most of the other encryption algorithms [6]
relevant nowadays, we always find permutation operators as essential building
blocks inside. In this contribution we will introduce key-dependent permutation
operators of provably excellent quality inspired by chaotic Kolmogorov flows [1,
4]. From chaotic systems theory it is known [3] that the class of Kolmogorov flows
exhibits the highest degree of instability among all dynamical systems. As will
be demonstrated in the sequel, these outstanding properties make them a per-
fect inspiration for developing a novel class of strong cryptographic permutation
operators.
must partition the side length n. Furthermore let the quantities qi be defined by
qi = nni and let Ni specified by N1 = 0 and Ni = Ni−1 + ni−1 denote the left
border of the vertical strip that contains the point (x, y) to transform. Then the
discrete Kolmogorov system Tn,δ will move the point (x, y) ∈ [Ni , Ni +ni )×[0, n)
to the position
Tn,δ (x, y) = (qi (x − Ni ) + (y mod qi ), (y div qi ) + Ni ). (2)
In the final paper we will prove the following theorem:
Theorem 1. Let the side-length n = pm be given as integral power of a prime
p. Then the discrete Kolmogorov system Tn,δr as defined in equation 2 fulfills
the properties of ergodicity, exponential divergence and mixing provided that at
least 4m iterations are performed and lists δr used in every step r are chosen
independently and at random.
Based on this theorem it is well justified to claim that the permutation op-
erator developed in this contribution is indeed an excellent key-dependent per-
mutation operator for cryptographic applications.
References
1. V.I. Arnold and A. Avez. Ergodic Problems of Classical Mechanics. W.A. Benjamin,
New York, 1968.
2. J. Daemen and V. Rijmen. AES proposal: Rijndael. In First Advanced Encryption
Standard (AES) Conference, Ventura, CA, 1998.
3. S. Goldstein, B. Misra, and M. Courbage. On intrinsic randomness of dynamical
systems. Journal of Statistical Physics, 25(1):111–126, 1981.
4. Jürgen Moser. Stable and Random Motions in Dynamical Systems. Princeton Uni-
versity Press, Princeton, 1973.
5. US Department of Commerce National Bureau of Standards. Data encryption stan-
dard. NBS FIPS PUB 46-1, 1977.
6. Bruce Schneier. Applied Cryptography. Addison-Wesley, 1996.
Fault Cryptanalysis of ElGamal and Related
Signature Schemes
1 Introduction
In this paper we focus on six different versions of ElGamal and related DSA
and Schnorr signature schemes. It will be shown that, choosing algorithm to a
particular application has huge impact to security, which will decrease when one
is able to induct errors during computations. Our attack assumes that signing
purposes are implemented in tamper-proof device and one is able to induct
random bit-flip error to private key a stored in this device. We also assume
that each error affects only one bit and the probability that i-th bit of key a is
erroneous equals 1/n.
In our paper we present different signature schemes and their weakness
against fault analysis. We also present methods that improve some of those
schemes to remain secure. Those observations can be useful when modifying or
creating new signature schemes.
The scheme mentioned above is susceptible to fault cryptanalysis and one can
recover the key a with probability 1/2 after nlog2n erroneous signings. Complex-
ity of this attack is O(n2 logn). This can be easily proved.
Similar analysis can be done with respect to other versions of ElGamal
schemes. Those versions differs only in step 3 of signing which can have one
of the following forms:
3a. s = a−1 (h(m) − kr)
3b. s = ar + kh(m)
3c. s = ah(m) + kr
3d. s = a−1 (r − kh(m))
3e. s = k −1 (r − ah(m))
As we show in our paper versions 1 and 4 are immune to fault cryptanalysis
of the discussed type. On the other hand other versions of ElGamal scheme
are susceptible to fault cryptanalysis and can be broken with probability 1/2 in
nlog2n attempts. We will also present and examine modifications that allow to
increase cryptanalysis complexity and thus improve security.
In our paper we also study DSA and Schnorr signature schemes. We show that
DSA scheme cannot be broken using the same technique as in case of ElGammal
thanks to mathematical modifications that influence key preparation and signing
processes. Unfortunately there are other methods how one can break DSA. Those
methods were presented in [2].
We have also examined Schnorr signing scheme. In the article we present
how one can recover secret integer a from tamper-proof device without solving
discreet logarithm problem or reversing one-way function. This can be done
despite the Schnorr scheme use the same mathematical backgrounds for key
generation as DSA scheme. Moreover such attack requires less mathematical
computations and can be carried out easier than in case of ElGamal scheme.
References
1. Boneh D., DeMillo R.A., Lipton R.J., On the Importance of Checking Crypto-
graphic Protocols for Faults, Advances in Cryptology - EUROCRYPT’97, LNCS
1233, pp.37-51, Springer-Verlag, 1997
2. Christophe Giraud and Erik W. Knudsen, Fault Attacks on Signature Schemes,
ACISP 2004, LNCS 3108, pp. 478491, Springer-Verlag, 2004.
Complexity-Theoretical Approaches to
the Design and Analysis of
Cryptographical Boolean Functions
Departamento de Informática
Universidad de Oviedo
{cobas,brugos}@epsig.uniovi.es
1 Introduction
Boolean functions are at the core of the design of every symmetric cipher algo-
rithm, be it stream or block type. There is a wide body of theory supporting the
design choices behind many of the standard algorithms in wide use [4].
However, in many respects the design of cryptographic-quality boolean func-
tions still depends on many ad-hoc constraints. Efficiency and compactness con-
siderations leave very little room for choice when other algorithm parameters
(key schedule, number of rounds, etc.) determine too stringently the properties
we need for the required S-boxes or feedback functions. Choosing at random is
possible if efficiency is neglected [1].
Although complexity theory lies deeply at the roots of cryptographic strength,
it seems too often neglected in the design and analysis of algorithms, in which
algebraic approaches are favored. However, measures of complexity have proven
useful for detecting weaknesses in preliminary phases of design [3].
In this paper we survey the usual criteria that apply to the design of boolean
functions for stream and block ciphers, proposing a complexity-theoretic frame-
work in which measures of complexity provide a test bench for nonlinear elements
of cryptographic algorithms in preliminary analysis.
– being balanced
– propagation criteria of various orders
– being bent
– avalanche criteria
The relation among all these criteria and many others have been thoroughly
studied and still are subject of intensive research [6, 5]. Their algebraic character
is remarkable.
3 Complexity-theoretic approaches
With a complexity-theoretic point of view, complexity measures can be applied
to test how good a non-linear boolean function could perform. Complexity mea-
sures that have been proposed, with and without cryptographic purposes in
mind, have been reviewed in [2]. Application of these measures can be done in
several ways:
4 A test bench
In this presentation we propose 4 and 5, together with the other well-established
techniques as procedures for assessing the suitability of boolean functions for
a given cryptographic purpose. In the framework of a reduced-round version of
the well-known block algorithms AES and Serpent, and a toy, proof-of-concept
stream cipher algorithm, we show how to apply this method to select the best
possible choice for the non-linear elements involved.
References
1. C. M. Adams and S. E. Tavares. Good s-boxes are easy to find. In Proceedings
of the 9th Annual International Cryptology Conference on Advances in Cryptology,
pages 612–615. Springer-Verlag, 1990.
2. J. D. Gonzalez-Cobas and J. A. Lopez-Brugos. A complexity-theoretic approach to
the design of good measures of cryptographic strength. In International Workshop
on Computer Aided Systems Theory, LNCS, volume 8, 2001.
3. J. L. Massey. Shift-register synthesis and BCH decoding. IEEE Transactions on
Information Theory, IT-15:122–127, 1969.
4. W. Meier and O. Staffelbach. Nonlinearity criteria for cryptographic functions. In
Proceedings of the workshop on the theory and application of cryptographic tech-
niques on Advances in cryptology, pages 549–562. Springer-Verlag New York, Inc.,
1990.
5. J. Seberry, X.-M. Zhang, and Y. Zheng. Structures of cryptographic functions
with strong avalanche characteristics (extended abstract). In Proceedings of the
4th International Conference on the Theory and Applications of Cryptology, pages
119–132. Springer-Verlag, 1995.
6. Y. Zheng and X.-M. Zhang. On relationships among avalanche, nonlinearity and
correlation immunity.
! " # $
" """ "
% &" " # $' Æ
"
" (
)
*+,
"
Æ - .
.
# $
"
) #$
" " ) / #/$'
'
" ) "
Æ
"'
"
"
" !
*0,
) *1,
" "
" *2, !
3
"'
4 "
""' ""' " '
*5, (
'
" )
)
6" 3'
3 "
4 "
"' *7,'
Æ "
' 8 9 #
$
'
/
"
"
#
$ ! "
"
8
)
" " % "'
¼
!
" "
#$
%&&'(&')(&'(&
"
8 """
"
" 9 # $
:
5
9 5
1
9
5
1
'
3
)
"" ' # $
/
;
'
" # 9 5
1
$'
# 9 5
1
$
"
# 9 5
1
$' "
<
9 #
$
9 5
1
'
"
9 5
1
' 9
"
'
"" "
"
< 9 ## $ "
$
# 9 5
1
9 5
1
$
/
"
'
9 :
"
=
/<
9 :'
5
1
< #
5
1
# $ 9
$ 9 #
$ "
*
"
+ , +- Æ
.
(/0!
"
#
! *
#
!
% #1!
* 2
3&&4 56(& 2)6
7
7
8 * 8- .
(90!
"
"
:! (
!34 354 (5)
79! ; , -
**
**!
*"
!
:!
!34 3&&&4 5(&
' %!"0
;! 9< -
90!
"
*!
=
(
* 8;( 35)4 5(&'
) %!"
;! ,"
8- >0
!! 2#(
*
?
(
90!
"
3564 () 26
6 #
"
#- 2(
" *
(
@
A1!
3554 5('& 2)6
½
¾
½
¾
!"#
$$ %%& '
!
!
! " !
# ! "
$
! " %
& ' ( )* !
$#$# +
,-)%
!
. "
!
!
!
" + " !
" #" "
!
/
"
"
" "
!
!
+
0
! " "
"
! ! /
123 !
#
!
"
0
$
/
" &
' 0
" 0 "
"
/
+
"
/ " 0
/
!
#
¼
(
#"
!)
#
*%%$+%$,+%$+%
+ !
"
! !
143 ,
" 5
163 "
"
!
!
! " #
!
Æ
7 .
8
99
#"
/
0
" " 0 "8
¯
8
"
0
¯
8
"
0
¯
8
"
0
#
0
)
" &
' 0
" ! ,
! !
*# - . (
#
#
/ "
!0#
Dieter Gollmann
TU Hamburg-Harburg
diego@tu-harburg.de
Extended Abstract
– Cryptography is “perfect”. The adversary does not try to exploit any weak-
ness in the underlying cryptographic algorithms but only algebraic properties
of cryptographic operators and interactions between protocol messages.
– The adversary can observe and manipulate all messages exchanged in a pro-
tocol run and can itself start protocol runs.
The second assumption was already stated by Needham and Schroeder [7]:
References
1. B. Aboba, L. Blunk, J. Vollbrecht, J. Carlson, and H. Levkowetz. Extensible Au-
thentication Protocol (EAP), June 2004. RFC 3775.
2. Tuomas Aura, Michael Roe, and Jari Arkko. Security of Internet location manage-
ment. In Proceedings of the 18th Annual Computer Security Applications Confer-
ence, pages 78–87, December 2002.
3. Danny Dolev and Andrew C. Yao. On the security of public key protocols. IEEE
Transactions on Information Theory, IT-29(2):198–208, March 1983.
4. Dieter Gollmann. Authentication by correspondence. IEEE Journal on Selected
Areas in Communications, 21(1):88–95, January 2003.
5. D. Johnson, C. Perkins, and J. Arkko. Mobility Support in IPv6, June 2004. RFC
3775.
6. Alfred J. Menezes, Paul C. van Oorschot, and Scott A. Vanstone. Handbook of
Applied Cryptography. CRC Press, Boca Raton, FA, 1997.
7. Roger M. Needham and Michael D. Schroeder. Using encryption for authentication
in large networks of computers. Communications of the ACM, 21:993–999, 1978.
8. Harald Vogt. Integrity preservation for communication in sensor networks. Technical
Report 434, ETH Zürich, February 2004.
Pattern Recognition in AVHRR Images by Means of
Hibryd and Neuro-Fuzzy Systems
Jose Antonio Piedra, Francisco Guindos, Alberto Molina, and Manuel Canton
Extended Abstract
The problem is to study methods for the automatic interpretation of ocean satellite
images to recognize meso and macroscalar ocean structures. The difficulty of the
image analysis and understanding problem for satellite data is due, in large part, to the
lack of a precise mathematical description of the observed structures and to their
variability. This work proposes an approximation that show the results obtained with
an automatic interpretation system for AVHRR (Advanced Very High Resolution
Radiometer) infrared images, developed to detect and to label oceanic structures. This
automatic interpretation system contains mainly a set of different classifier, that allow
to compare the results. The classifiers use functional and operational principles de-
rived from the ANNs (Artificial Neural Nets), Rule-Based Knowledge Systems and
Fuzzy System.
Fig. 1 depicts the overall structure of the system that has been developed. In a first
step, the raw image is processed by means of algorithms such as radiometric correc-
tion, map projection and land masking. The second step aims to detect clouds pixels
that are opaque to radiance data measured in the AVHRR infrared and visible
scenes[1]. The following task is the segmentation that will divide the whole image in
regions. The nature of ocean dynamics makes very difficult this process that is never-
theless fundamental, so we’ve designed an iterative knowledge-driven method to
perform this part of the process pipeline [2]. The next task is the features or descrip-
tors selection, which consists of selecting an optimal or sub-optimal feature subset
from a set of candidate features. The most common framework for features selection
is to define some criteria for measuring the goodness of a set of features [3], and then
use a search algorithm to find an optimal or sub-optimal set of features [4]. We have
used Bayesian networks for features selection in the recognition of oceanic structures
in satellite images [5][6]. In the last step, each region produced in the segmentation is
analyzed and, if the recognition is positive, it is labeled with the identifier of the
matching structure. The structures of interest in the Canary Islands zone as defined in
[7] are: coastal upwelling, warm eddies, cold eddies and island wakes.
Fig. 1 Structure of the ocean feature recognition system.
We have explored the use of Bayesian networks as a mechanism for feature selec-
tion in a system for automatic recognition of ocean satellite images. The use of
Bayesian networks has provided benefits with respect to SPE, not only in the reduc-
tion of relevant features, but also in discovering the structure of the knowledge, in
terms of the conditional independence relations among the variables. In future works
we plan to improve the accuracy rate of the system including more variables. Fur-
thermore, we expect to use models to avoid the discretisation of the continuous fea-
tures when learning Bayesian networks.
On the other hand, obtained results aren´t expected for classification process by
hybrid and neuro-fuzzy system. They don´t correspond with results of other cassifiers
(EPS, GES and Bayesian Network). The main problem is the number of generated
rules, which is excessive. In future works we plan to improve the accuracy rate of the
hybrid and neuro-fuzzy systems including more variables, tunning training parame-
ters, optimizing the obtained structure and automating the construction of neurofuzzy
system.
Finally, the structure of the automatic interpretation system that has been intro-
duced, it is a complex and independent structure for the following tasks:
1. Iterative segmentation-recognition cycle is made by means of graphic expert sys-
tem.
2. Feature selection and validation of knowledge is done by bayesian learning
3. Multiple classification allows to find different interpretations of the existing
knowledge to validate the knowledge of the classifiers.
References
1. Torres J.A., Guindos F., Peralta M. and Cantón M.: An Automatic Cloud-Masking System
Using Backpro Neural Nets for AVHRR Scenes, IEEE Transactions On Geoscience and
Remote Sensing, vol. 41, nº 4, (2003) 826–831.
2. Guindos F., Piedra J.A. and Cantón M.: Ocean features recognition in AVHRR images by
means of bayesian net and expert system, 3rd International Workshop on Pattern Recogni-
tion in Remote Sensing, Kingston University, United Kingdom. August (2004).
3. Yi Lu Murphey and Hong Guo: Automatic Feature Selection – a hybrid statistical approach.
IEEE, (2000) 382–385.
4. Langley P. and Sage S.: Induction of selective Bayesian classifiers. In Proceedings of the
Tenth Conference on Uncertainty in Artificial Intelligence, Seattle. Morgan-Kaufmann,
(1994) 399–406.
5. Bruzzone L.: Classification of remote-sensing images by using the Bayes rule for minimun
cost. IEEE Int. Geoscience and Remote Sensing Symposium, (1998) 1778–1780.
6. Yamagata Y. and Oguma H.: Bayesian feature selection for classifying multi-temporal SAR
and TM Data. IEEE, (1997) 978–980.
7. Arístegui J., Sangrá P., Hernández-León S., Cantón M., Hernández-Guerra A. and Kerling
J.L.: Island-induced eddies in the Canary Islands, Deep-Sea Research, nº 41, (1994) 1509–
1525.
8. Zadeh L.A.: Fuzzy Logic. IEEE Computer, vol. 21 , issue 4 , April (1988) 83–93.
9. Roger Jang: ANFIS Adaptative Network based Fuzzy Inference System. IEEE Transactions
On Systems, Man and Cybernetics, vol 23, nº 3. May/June (1993).
10. Nauck D. and Kruse R.: A neuro-fuzzy approach to obtain interpretable fuzzy systems for
function approximation. Fuzzy Systems Proceedings, IEEE World Congress on Computa-
tional Intelligence., vol 2 , May (1998) 1106 –1111.
From Projective Virtual Reality to the Virtual Human:
Virtual Worlds become realistic
Jürgen Rossmann
Extended Abstract
Latest experiences in industrial and space applications have shown, that the symbiosis
of virtual reality techniques with versatile robot control techniques can break new
grounds in the field of virtual reality applications as well as in the field of robotics. In
this context the new paradigm of Projective Virtual Reality is presented as a
sophisticated new approach to the control and supervision of robotic and automation
systems.
The idea behind Projective Virtual Reality is to "project" actions that are carried out
by users in the virtual world into the real world, primarily by means of robots but also
by other means of automation. Robot control technology thus provides the user in the
virtual reality with a "prolonged arm" into the physical environment, paving the way
for a new quality of a user-friendly man machine interface for automation
applications. The paper provides the background of Projective Virtual Reality which
is based on latest results in the fields of task planning, world modeling, cooperative
robot control and sensor-based robot control. Building on these basic ideas, new
research directions related to simulating the “virtual human” will be presented. The
implemented approach of the real-time dynamic simulation of an anthropomorphic
kinematic chain will be demonstrated. The implemented virtual human in the
applications presented not only serves as an ergonomic “role model” in a digital
factory environment, but it is also flexible enough – and easy enough to use – to be
used for general scientific applications ranging from the control of anthropomorphic
agent, the display of emotions by body language and various medical applications.
Fig. 2. The dynamic simulation of an astronaut on board the International Space Station
The first step to make virtual worlds “feel” realistic was to simulate the underlying
physics. We are currently pursuing the next step which is to populate these “physical
virtual worlds” with human shaped avatars to further increase user-acceptance. Will
we be able to simulate a virtual society some time soon?
Extended Abstract
The traditional method of auscultating heart sounds is still the first basic analysis tool
that is applied in the evaluation of the functional state of the heart, as well as the first
indicator used to submit the patient to a cardiologist. A well-trained medical doctor can
reach a final diagnosis immediately after an exploration and a conscientious ausculta-
tion. During the last 15 years, heart auscultation has undergone the spectacular advance
of other high-technology diagnosis techniques such as the doppler-ultrasound transtho-
racic echocardiogram (TTE). Nevertheless, the needed high-technology devices can be
exclusively found at high-level hospitals, while it is absent in most cardiological ambu-
latory services as well at Primary Care Centers for adult and pediatric population.
?
This work was supported by the Spanish MCyT under project MEDICI (TIC2003-09400-C04-
01) and Murcia Regional Government under project PB/46/FS/02
Nevertheless, in recent years, the medical community has shown a renewed interest
in cardiological auscultation and PCG. This renaissance of PCG is motivated by the
confluence of a diversity of factors: a) the present situation of overutilization of high
technology devices (such as TTE), which are restricted to a few Cardiology units, b)
the demonstrated usefulness of cardiac auscultation in clinical prediagnosis, especially
in primary attention, since some pathologies can be detected by this process. Further-
more, many pathological conditions of the cardiovascular system cause murmurs and
alterations in the sounds of the heart, even before they were detected as symptoms by
means of other more advanced techniques [3].
In this work, a system for the storage, visualization, automatic detection and di-
agnosis of alterations of the heart function, based on the analysis of the PCG signal, is
proposed. Automatic detection and diagnosis requires a previous characterization of the
PCG morphological properties, in order to proceed with the identification. The process-
ing of the PCG signal is base on a hierarchical structure that provides a methodology
for the analysis and interpretation of the phonocardiographic signal, with a scheme sim-
ilar to that followed by the physician during auscultation. This structure is based in the
abstraction of signals and organized in different levels, each one with different blocks
aimed to specific tasks, which perform the complete processing from the acquisition of
the phonocardiographic record to the final diagnosis. Besides the abstraction levels [4],
in each level several specialists exist, which are processing blocks that perform analysis
of the associated signals or extract some kind of information. Each specialist is defined
with input data, grouping operations, and output data, which will be available to the rest
of specialists of the same level and higher. The lowest level of the hierarchy performs
the most intensive computations, whereas as the levels are higher, the information is be-
ing abstracted progressively. Examples of methodologies like the proposed in this paper
have been applied successfully to other biomedical environments, like the detection of
myocardial ischaemia [5]. The proposed hierarchy is organized in four levels, each one
associated to the main signals: envelopes, events, sounds, and diagnosis.
The proposed system pursues a double goal. On the one hand that of serving as ob-
jective element for prediagnosis of alterations of the heart function in primary attention.
On the other hand, the second goal is focused on the exploitation of training capabili-
ties of the system. To achieve these goals, a multimedia database is provided in which
the PCG signal and the processing results are stored together with the information of
patient’s clinical history. In this multimedia database, different metaphors for the PCG
visualization are offered, as well as the corresponding TTE video, thus making possible
for the user to associate PCG sounds with real heart images. Finally, an important effort
has been made in providing the medical stuff with an adequate system for the acqui-
sition and electronic storage of the PCG signal. In this sense, an ubicuous acquisition
system, based on PDA, has been developed. This PDA acquisition systems allows the
medical stuff to carry out the auscultation process, and at the same time, allows them to
add annotations to acquired the PCG signal.
References
1. Mangione, S., Nieman, L.: Cardiac auscultatory skills of internal medicine and family practice
traineers. Journal of the American Medicla Association 278 (1997) 717–722
2. Gaskin, P., Owens, S., Talner, N., Sanders, S., Li, J.: Clinical auscultation skills in pediatrics
residents. Pediatrics 105 (2000) 1184–1187
3. Lukkarinen, S., Noponen, A., Angerla, A., Sikiö, K., Sepponen, R.: A recording stethoscope
in study of heart sounds of children. Medical & Bilogical Engineering & Computing 34 (1996)
97–98
4. Milios, E., Nawab, S.: Signal abstractions in signal processing software. IEEE Trans. on
Acoustics, Speech, and Signal Processing 37 (1989) 913–928
5. Ruiz, R., Barro, S., Presedo, J., Palacios, F., Vila, J.: Abstraction of information in elec-
trocardiographic monitoring for the detection of myocardial ischaemia. In: Proc. of the V
International Symposium on Biomedical Engineering. (1994) 291–294
Dactylology Language Translator from Video to Text
Abstract. In this paper an innovative system has been created, which detects
the movement of the hand and implements a converter of dactylology language
to text, segmenting an input video and recognizing the frames selected, by
means of an algorithm of a movement detector system. Once, it established the
correct images, these have been parameterizated and recognized with a decision
tree classifier and a network neuronal (NN), from the characteristics of the
different classes of the dactylology alphabet, with a success of the 93.9%. The
reached times in these tests show their possible application in real time.
1 Introduction
Since the principle of the times, the human species has gone feeling the need to have a
greater quality of life. Along the history, the technological development has achieved
an evolution that contributes clearly to create a welfare society as we find it
nowadays. In spite of the conscience of global social integration, it exists minorities
represented by collective or groups that cannot opt to an improvement of the quality
of life, since they do not possess the same options neither opportunities that the
remainder of the citizens. The present associate-cultural tendency produces an
accelerating continuous in the need of routine interaction of the users with diverse
systems, which take advantage to try to break the barriers that impede the integration
of these groups [1]. For this reason, the motivation of this work comes given by the
application of a system to recognize the movement of the hand (video) to text, with
the purpose to be able to break the barrier of language among the people with auditory
decrease and the remainder of the society.
Captured video
Web cam
2 Movement detection
To carry out this conversion we need a device sensor (web camera, video camera
more acquiring card, digital or even, a mobile telephone with camera integrated) for
receiving the images that contain the gesticulation.
To obtain a capture of the correct gesture has been utilized a uniform background,
for the acquisition of the video samples, complying a series of environmental
conditions as the brightness, the position of the hand on the background, etc. Also,
some own parameters of the capture are established among them, we emphasize the
shutter speed of the input device, which has been acquired to 30 frames by second.
The extraction of the frame (of the input video sequence), pertaining to a definite
gesture, is carried out by the monitoring of the movement of the hand, utilizing a
movement detector system (SDM), designed to consider the dynamic conditions of
the gesticulation.
In the above algorithm (SDM), as the first step is carried out a test of the video
each 4 frames, this signifies that the sequence is analyzed each 0.13 seconds. On this
way, we will be grasped with security all the possible gestures carried out. From each
frame extracted, some peripheral parameters are calculated that give an idea of the
position of the hand in that determined instant. These data are stored to do a
comparison with the same data but of the following frame. Thus, we can have an idea
of the movement carried out by the hand, in the transition between the two captures
(elapsed 0.13 seconds). For it, two types of movements are defined, depending on the
comparison of the peripheral distances calculated with respect to some thresholds
created for the operation of the SDM:
When a static gesture is detected, with respect to the previous frame, is selected in
order to extract its characteristics for its subsequent classification. The remainder of
the frames are ruled out as transition or stable frames due the static position of the
hand with the time. It is possible to capture the same gesture more than one time,
since the hand can fluctuate of natural form or, even, the gesticulation can carry out
repercussions due to the inexperience of the use of the language of signs on the part of
the person that makes the shot. For this reason, the segmented gestures have been
studied to avoid a repetition of the same, ruling out those whose characteristics are
very similar. At the same time, it will serve to do a classification that summons to the
gestures to some specific groups depending on its morphology.
Variability
Light
Movement
Stable
0.13 0.26 0.39 0.52 0.65 Seconds
Achieving it, some general characteristics are extracted, which gives an idea of the
dimensions, area and solidity of the gesture. Subsequently, perimeter recognition is
done to find more specific parameters, as the number of fingers, inclination of the
hand, etc. Attending to these characteristics calculated will proceed to a classification
in which the gestures with similar properties establish groups that define the type of
gesture. The creation of the groups is carried out to obtain an organization of the
alphabet in the way of tree; this causes a faster recognition of the segmented frames.
Analyzing the alphabet are found gestures, whose similarity is evident, as it is the
case of the letters ‘s’ and the ‘o’. Solving it, some more specific parameters are
extracted, which help to recognize the relation among the similar gestures.
To extract the characteristics of the gesture processed [2][3] and to assign them for
a classification, a decision tree classifier is created gestures for the all the gestures of
the alphabet. In some cases the sub-classification of the alphabet can go to the
creation of the exclusive groups for a single gesture, in such case, a direct answer is
obtained, which do not need that the gesture be processed to take it like input of the
recogniser.
The characteristic parameters of each gesture segmented are stored in sequential
form, to carry out an exhaustive comparison of these parameters, so much the generic
as the specific parameters. On this way, we can know that gestures must be
discriminated and which ones must be recognized. Therefore, this system is able of
detecting the characteristics of two similar consecutive gestures, in this case, the last
processed frame will be ruled out. If they are different, the two frames will be
analyzed since they are taken like valid and different samples.
The gestures of the input video, that are not recognized directly, need a pre-
processing in order to prepare the data to the input of the neuronal networks (NN) [4],
which form the classifier based on the groups of letters, that are established in the
characteristics extraction.
For the analysis and evaluation of this system are needed two databases. A
dynamic database, that is included for short videos where are represented the
variability of the gestures of the hand, and the form of the movement for its
segmentation is verified. And another one is static [5], which serves for training and
test the NN [4]. In the following figure, we can see one sample of each gesture.
4 Conclusions
To conclude, in this paper have created an innovative system that detects the
movement of the hand for the Spanish dactylology language, that is capable of
segmenting an input video and to recognize the frames selected by a classifier based
on decision tree of the dactylology alphabet and a NN [4]. This system reaches an
average success rate of 93.9%, with a low computation cost in all their steps, which is
possible suitable for real time applications (compiling the programs implemented
from MATLAB to C language), therefore, this system permits to do an online
translation of the gesticulation.
5. Acknowledgement
This work has been developed and agrees by research Project UNI2003/26 from
University of Las Palmas de Gran Canaria (Spain).
References
1. Gascón-Ricao A.: http://www.ucm.es/info/civil/bardecom/docs/signos.pdf. ¿Señas o
signos?: Evolución histórica. [Active on October 2004]
2. O’Gorman L., Kasturi R.: Document image analysis. IEEE Computer Society Press, 1996
3. “Image acquisition toolbox”, User’s Guide, Version 5, The Math Works, INC, (2004)
4. Bishop C.M.: Neuronal network for pattern recognition. Ed.Oxford University Press,
(1994)
5. Travieso C.M., Alonso J.B., Ferrer M.A.: Sign Language to text by SVM. Proceedings of
Seventh International Symposium on Signal Processing and its Applications, vol. 2,
(2003) 439-442.
Automatic Braille Character Recognition: BrailLector
1. Introduction
Nowadays, lack or problem of vision has been an important obstacle to access to
printed contents and to the information society. For this reason, some people have
tried to achieve that blind people are able to access to the printed culture, for example
Valentin Haüy and Luis Braille [1] who have understood the importance of a
communication code. Globally, an estimated 40 to 45 million people are blind and
135 million have low vision according to the World Health Organization (WHO) [2]
and this number grows every year.
A Braille Optical Character Recognizer is interesting due to the following reasons;
Since the most part of Braille books are written with two kinds of points, we will
have to distinguish between each one. One kind is like “mountains” and the other is as
little “valleys”, hence the finger of the blind person who read the text only detects the
protrusions, while depressions are points written to be read from the other side of the
sheet. If we are able to distinguish between each point we will have a big advantage,
we will be able to recognize the two sides with only one scan.
Fig. 1. Protrusions and Depressions on a Braille sheet
The structure of this paper is the next: In the next text we will describe the
characteristics of the database created. After it will be described image processing
techniques used for dots detection and recovering and Conclusions closes this
extended abstract.
The mechanism used for image acquisition has been a flat-bed scanner because it is
a cheap alternative which can be used for so many other applications and it is easy
and quick to use.
The different steps of the scanned image for its translation are;
4. After mesh building, all valid Braille positions are known. Now we fit this image
on the dots image after thresholding in order to check in detail those positions
where dots were not found. In this point lays the intelligence of the global
system, only potential positions of dots are checked; Original dots will be
recovered (they belong to correct Braille positions) and false dots will be
discriminated as they are out of the valid places for a Braille dot.
The final image with the detected dots is analyzed and text is segmented in rows
and characters. Every character will be converted into a binary number according
to the active dots.
3. Conclusions
Finally, in this paper we have developed an automatic system for translating Braille
text to normal text or voice. The global algorithm is very fast and robust. For
achieving this system, dynamic thresholding and adaptive Braille grid has been used,
adding some intelligence to the global process and making it able to detect dots in
both sides of the document with only one scan. This process has an efficiency of
99.9% and it takes only 26 secs to translate a double-sided document improving all
the main references found in the bibliography [4],[5],[6].
4. Acknowledgement
This work has been developed and agrees by research Project UNI2003/26 from
University of Las Palmas de Gran Canaria (Spain).
References
1. http://www.asac.org.ar/Publi/bra_b001.htm Active on October 1st 2004
2. http://www.who.int Active on October 1st 2004
3. Dubus J., Benjelloun M., Devlaminck V., Wauquier F., Altmayer P.: Image processing
techniques to perform an autonomous system to translate relief Braille into black-ink,
called: Lectobraille. In Proceedings of the Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, (1988) 1584–1585.
4. Wong L., Abdulla W., Hussman S.: A Software Algorithm Prototype for Optical
Recognition of Embossed Braille. The University of Auckland, New Zealand, (Technical
Report) (2004).
5. Mennens J.,. Tichelen L. V, Francois G., Engelen J.: Optical recognition of braille writing
using standard equipment. IEEE Transactions on Rehabilitation Engineering, 2(4) (1994)
207– 212.
6. Ng C., Ng V., Lau Y.: Regular feature extraction for recognition of Braille. In Proceedings
of Third International Conference on Computational Intelligence and Multimedia
Applications, (1999) 302-306.
Retinal Based Authentication via Distributed Web
Application
1 Introduction
Reliable authentication of people has long been an interesting goal, becoming more im-
portant as the need of security grows, so that access to a reliable personal identification
infrastructure is an essential tool in many situations (airport security controls, all kinds
of password-based access controls, ...). Conventional methods of identification based
on possession of ID cards or exclusive knowledge are not altogether reliable. ID cards
can be lost, forged or misplaced; passwords can be forgotten or compromised. A solu-
tion to that problems has been found in the biometric based authentication technologies.
A biometric system is a pattern recognition system that establishes the authenticity of
a specific physiological or behavioral characteristic possessed by a user. Identification
can be in the form of verification, authenticating a claimed identity, or recognition, de-
termining the identity of a person from a database of known persons (determining who
a person is without knowledge of his/her name). Many authentication technologies can
be found in the literature, with some of them already implemented in commercial au-
thentication packages [1–3]. But today most of the efforts in authentication systems
tend to develop more secure environments, where it is harder, or ideally, impossible,
to create a copy of the properties used by the system discriminate between authorized
individuals and unauthorized ones, so that an impostor could be accepted by the bio-
metric system as a true sample. In that sense, the method employed by our system [4]
uses for authentication biometric parameter the blood vessel pattern in the retina of the
eye.
In this paper, a distributed client-server framework which allows for personal au-
thentication is outlined. From the design stage, where design patterns are the funda-
mental tool, to the implementation (server,communication channels and clients) are de-
scribed. The main goal of this work is to build a robust system which allow for the
automatic personal authentication, which would allow or deny the access to critical
resources of the system.
Web/application server
Authentication server
HTTPS
Acceptance/
rejection
Users database
The more important tools employed for designing the system has been the design
patterns [6] and the UML [7, 8], which allow for a more robust and reliable software
applications.
A client application initiates a connection to authenticate an individual, and if au-
thentication is successful, resources can be accessed. The graphical user interface for
the interaction with data (like the depicted in Figure 2) is obtained through the web
application.
The overall operation flow at the client-side can be described as follow (see also
Figure 1):
– An individual makes HTTP request to web server first. The web server will handle
the request and will communicate with the authentication server to perform the
authentication process.
– The authentication server will access the database and will allow or deny the access
to the requested resource. Upon request, procedure execution will be invoked (using
XML-RPC) based on the data sent by the client.
– Once the procedure ends, the related information is sent to the client and displayed
in the browser window, allowing or denying access depending on the result of the
authentication process.
Fig. 2. Screen from the client program-application with retinographies, which are the biometric
authentication parameter (a) and obtained authentication parameters (b).
3 Communication technology
To assure confidentially and reliability in the communications, the most widely used
secure Web encryption has been employed: the Secure HyperText Transport Protocol
(HTTPS) on the Secure Socket Layer (SSL) version 3.0 with 40 and 128 bit RC4 en-
cryption. The SSL [10] protocol has become most widely used method for encrypting
and authenticating Web communications.
Our approach to deal with the heterogeneity of the clients focuses on documents or
data and transforms that visualizes the data to fit the computing capabilities of differ-
ent users, while preserving semantics of data. This division of the underlying data and
the way it gets displayed mimics the Model-View-Controler (MVC), a well known and
frequently used design pattern to develop interactive applications with flexible human-
computer interfaces. This would allow for a future implementation of mobile authenti-
cation clients, for example.
XML has been chosen as the medium to represent the data sent to the server for
the authentication. XML is a markup language for documents containing structured
information. Once the data is written in XML, as associated XSL document can be
written to define the way it gets displayed. The XML document is platform-independent
and thus corresponds to the model, whereas the style sheet depends on the displaying
device, and corresponds to the view.
MVC separation of view and model offers the following advantages over conven-
tional application-centric environments:
Finally, using XML-RPC [11], procedure calls are wrapped in XML establishing
that way a simple pathway for calling authentication functions. XML-RPC is a quick-
and-easy way to make procedure calls over the Internet. It converts the procedure call
into XML document, sends it to a remote server using HTTP, and gets back the re-
sponse as XML. Other considered (although finally discarded) protocols where SOAP
and CORBA, both of them popular protocols for writing distributed, object-oriented ap-
plications, well-supported by many vendors, although they are much more complex than
XML-RPC,and require fairly sophisticated clients. In addition, XML-RPC has fewer
interoperability problems than them.
Acknowledgments
This paper has been partly funded by the Xunta de Galicia through the grant contract
PGIDIT03TIC10503PR
References
1. J.G. Daugman. Biometric personal identification system based on iris analysis. United States
Patent No.5,291.560, 1994.
2. J.Bigüin, C.Chollet, and G.Borgefors, editors. Proceedings of the 1st.International
Conference on Audio- and Video-Based Biometric Person Authentication, Crans-
Montana,Switzerland, March 1997.
3. R. Zunkel. Hand geometry based verification. In BIOMETRICS:Personal Identification in
Networked Society. Kluwert Academic Publishers, 1999.
4. C. Mariño, M.G. Penedo, and F. González. Retinal angiographies based authentication. Lec-
ture Notes in Computer Science, 2905:306–313, 2003.
5. Robert Orfali, Dan Harkey, and Jeri Edwards. Client/Server Survival Guide. John Wiley &
sons, 3rd edition, 1999.
6. E. Gamma, R. Helm, R. Johnson, and Vlissides J. Design Patterns, Elements of Reusable
Object-Oriented Software. Professional Computing Series. Addison-Wesley, 1995.
7. Mark Grand. Patterns in Java: a catalog of reusable design patterns illustrated with UML,
volume 1. New York,John Wiley & sons, 1998-1999.
8. Mark Grand. Patterns in Java: a catalog of reusable design patterns illustrated with UML,
volume 1. New York,John Wiley & sons, 1998-1999.
9. Java 2 sdk, standard edition version 1.3.1, plug-in installation notes.
http://java.sun.com/j2se/1.3/install-linux-sdk.html.
10. K.E.B. Hickman. The SSL protocol. http://www.netscape.com/newsref/ssl.html, December
1995.
11. Simon St.Laurent, Joe Johnston, and Edd Dumbill. Programming Web Services with XML-
RPC. O’Reilly, 2001.
Medial Axis Transform Computation Using
Shock Wavefront Detection
1 Introduction
The medial axis transform (MAT) or skeleton of a geometric figure is a set of
curves that approximate the local symmetry axis of this figure. The concept of
medial axis consists in the reduction of geometric shape information to the min-
imum. The first definition was given by Blum [1], who stated the analogy with a
prairie fire: the MAT is the set of points where the propagation wavefront initi-
ated from the shape boundary “intersect itself”. Despite the extreme simplicity
of this definition, the implementation in digital images without loss of important
properties is surprisingly difficult. For this reason there exists a large number of
algorithms and methods in the literature to address this problem. The relevance
of MATs was introduced initially by Blum mainly for visual perception analysis
and it is used especially for pattern recognition.
Formally there exist four definitions for medial axis transform, all of them
equivalent, described by Montanari in [2]. In the first of them the medial axis
transform is the prairie fire model proposed by Blum commented above where
the fire lines represent the propagation fronts. The second definition states that
the MAT is the set of ridges of the distance map constructed from the shape
boundary. In the third definition, one of the more common, the MAT is the set of
centers of the maximal discs of the figure, where a maximal disc is a cicumference
contained in the shape for which there exists no other circumference inside the
shape that contains it. The last model define the MAT as the set of points
that do not belong to any straight line segment connecting other points to their
respective closest boundary points.
The principal MAT properties are the connectivity, the preservation of the
main topological features of the shape, i.e. to not lose any main branch in the
skeleton and to not add irrelevant branches to it, and the Euclidean equidistance
of every point of the skeleton to two or more points on the boundary. Regarding
the discrete MAT, it is desirable to obtain one pixel thickness results to accom-
plish the narrowness property. Obviously in a discrete domain the equidistance
property is not possible to obtain, so it is necessary to relax this property to get
the most approximated result.
2 Method
Methods to construct MAT are well described in the literature [3–5]. In this
paper we propose a method to extract the MAT from a 2D shape, where the
distance map from the shape boundaries is computed using a propagation scheme
and then, the MAT is constructed detecting the points where the collision from
the wavefronts occurs.
The way of detecting the points where the wavefront intersection takes place
is carried out by a non propagation criterion. The discrete propagation fronts
initiated from the shape boundary, are implemented with a list of pixels at Eu-
clidean distance d from it, where d is an integer number, and these pixels are the
ones which have real distance values closer to d inside the shape. The wavefront
at distance d from the boundary will generate the wavefront at distance d + 1,
as long as there exist more inner non visited pixels. When a pixel can not be
propagated deeper into the shape, it is automatically labeled as member of the
skeleton. With this criterion the MAT obtained is not connected and has not
one pixel thickness.
The connectivity is not assured because two effects. The first one appears
because points which should be labeled as belonging to the MAT, are not de-
tected in cases where the propagation fronts coming from opposite directions do
not intersect at the same pixel but adjacent ones, as in figure 1(a). The second
effect happens when the shape becomes wider as illustrated in figure 1(b). The
first effect is corrected a posteriori using the distance map computed, connecting
the pixels of the computed MAT through the points with maxima values in the
distance map. The second effect is solved in the propagation process, by con-
necting the extremal points of the MAT computed in d (if any) with its nearest
point in the front d + 1. The narrowness property is easily achieved by a regular
d=3
d=2
d=2
d=1 d=1
d=0 d=0
(a) (b)
Fig. 1. Disconnection effects: wavefronts coming from opposite directions do not inter-
sect at the same pixel but adjacent ones (a), and aperture zones (b)
thinning which is highly efficient because only a small percentage of the image
has to be scanned, and only one thinning iteration is needed.
Acknowledgments
This work has been partially funded by an European grant FP6-507609 from
the European Commission, and has been partially supported by the Spanish
Ministry of Science and Technology and European Commission, co-funded grant
TIC-2001-38008-C02-01.
References
1. H. Blum, “A transformation for extracting new descriptors of shape,” in Proc.
Models for the Perception of Speech and Visual Form, Weiant Wathen-Dunn, Ed.,
Cambridge, MA, November 1967, pp. 362–380, MIT Press.
2. U. Montanari, “A method for obtaining skeletons using a quasi-euclidean distance,”
Journal of the Association for Computing Machinery, vol. 15, no. 4, pp. 600–624,
1968.
3. C.J. Ammann and A.G. Sartori-Angus, “Fast thinning algorithm for binary images,”
Image Vision Comput., vol. 3, no. 2, pp. 71–79, 1985.
4. R. Ogniewicz and M. Ilg, “Voronoi skeletons: Theory and applications,” in Proc.
IEEE Conf. Computer Vision and Pattern Recognition, CVPR, Los Alamitos, Cal-
ifornia, 15–18 1992, pp. 63–69, IEEE Computer Society.
5. R. Kimmel, D. Shaked, N. Kiryati, and A. M. Bruckstein, “Skeletonization via
distance maps and level sets,” Computer Vision and Image Understanding: CVIU,
vol. 62, no. 3, pp. 382–391, 1995.
6. B.H. Verwer, P.W. Verbeek, and S.T. Dekker, “An efficient uniform cost algorithm
applied to distance transforms,” IEEE Transactions on Pattern Analysis an Ma-
chine Intelligence, vol. 11(4), pp. 425–429, 1989.
Comparative Analysis of Calibration Methods
for a Static Camera
J. Isern González, J.D. Hernández Sosa, A.C. Domı́nguez Brito and A.M.
Naranjo Cabrera
Extended Abstract
References
1. M. Ahmed, E. Hemayed, and A. Farag. Neurocalibration: a neural network that can
tell camera calibration parameters. In IEEE International Conference on Computer
Vision, pages 463–468, September 1999.
2. X. Armangu, J. Salvi, and J. Batlle. A comparative review of camera calibrating
methods with accuracy evaluation. In Proceedings of 5th Ibero-American Sympo-
sium on Pattern Recognition, SIARP 2000, pages 183–194, September 2000.
3. J. Batista, H. Araújo, and A.T. de Almeida. Iterative multi-step explicit camera
calibration. In Proc. Sixth International Conference on Computer Vision ICCV98,
volume 15, pages 709–714, January 1998.
4. Olivier Faugeras. Three-dimensional computer vision: a geometric viewpoint. MIT
Press, 1993.
5. J. Isern Gonzlez. Estudio comparativo de mtodos de calibracin y autocalibracin
de cmaras. PhD thesis, University of Las Palmas de Gran Canaria, Las Palmas,
Spain, July 2003.
6. J. Heikkila and O. Silvén. A four-step camera calibration procedure with implicit
image correction. In Proc. IEEE Conference on Computer Vision and Pattern
Recognition, volume 1, pages 1106–1112, June 1997.
7. Jim Z.C. Lai. On the sensitivity of camera calibration. Image and Vision Com-
puting, 11(10):656–664, December 1993.
8. J. Salvi, X. Armangué, and J. Batlle. A comparative review of camera calibrating
methods with accuracy evaluation. Pattern Recognition, pages 1617–1635, June
2002.
9. Roger Y. Tsai. A versatile camera calibration technique for high-accuracy 3d
machin vision metrology using off-the shelf tv cameras and lenses. IEEE Journal
of Robotics and Automation, RA-3(4):323–344, August 1987.
10. Roger Y. Tsai. Synopsis of recent progress on camera calibration for 3d machine
vision. The Robotics Review, pages 147–159, 1989.
11. Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transac-
tios on Pattern Analysis and Machine Intelligence, 22(11):1330–1334, November
2000.
Cue Combination for Robust Real-Time
Multiple Face Detection at Different
Resolutions?
1 Introduction
The face detection problem, defined as: to determine any face -if any- in the
image returning the location and extent of each [7], seems to be solved, accord-
ing to some recent works [4] [6]. Particularly for video stream processing, these
approaches focus the problem in a monolithic fashion, forgetting elements that
the human system employs: temporal and contextual information, and cue com-
bination.
The work summarized in this abstract describes an approach for robust real-
time multiple face detection which combines different cues. The resulting ap-
proach achieves better detection rates for video stream processing and cheaper
processing costs than outstanding and public available face detection systems.
3 Results
For still images the lower boundary is the combination of the Viola-Jones based
detectors [3, 6]. If both detectors return a face in the same position, it is preferred
the frontal face detection as it is more precise. For still images the added value
of the approach is the likely eye detection for almost frontal views and the
combination of two Viola-Jones based detectors, see Figure 1.
Fig. 1. Detection examples from some CMU database [4] samples. The color indicates
the technique used: green means that the eyes were detected, yellow means that they
were not detected, and red containing a yellow rectangle means that the local context
detector was used. The images have been scaled down to fit the extended abstract size.
The size of the images are originally 814 × 820, 256 × 256, 611 × 467 and 159 × 160
respectively in the first row and in the second 500 × 500, 539 × 734, 336 × 484 and
258 × 218. Obviously for still images there are no detections based on color or tracking.
The benefits of our approach are evidenced in video stream processing. The
results achieved processing a set of 70 desktop sequences containing more than
35000 images of different individuals allow, for typical webcam resolutions 320 ×
240, multiple face detection in real-time, an average of 20 fps. The approach
provides better recognition rates than the OpenCV implementation of the Viola
frontal face detector (3 percentage points greater), faster, and with the added
value of eye detection for almost frontal faces. The approach is also suitable for
sequences with resolution changes, see Figure 2.
The temporal coherence applied to video stream processing not only speeds
up the process but also allows to despise most false detections which appear
when a still image is processed, see some false detections in Figure 1. In at
least 10 of the sequences some detections were non face patterns, and they were
correctly not assigned to any detection thread as the eyes were not found and
their position, or their color and size were not coherent with any active detection
thread. Or in the worst case, a non face detection was associated to a detection
thread, but the system observed soon an incoherence and decided to remove the
detection thread and wait for a new one, i.e. a new eye pair detection.
As a conclusion, the approach developed provides good and real-time per-
formance for multiple face detection at different resolutions, making the system
suitable for further processing in the field of perceptual user interfaces.
Fig. 2. Frames extracted from a video stream with 720 × 576 resolution. The color has
the same meaning than in Figure 1, but observe that the last frame depicts a blue
rectangle which means that tracking was used.
References
1. M. Castrillón Santana, F.M. Hernández Tejera, and J. Cabrera Gámez. Encara: real-
time detection of frontal faces. In International Conference on Image Processing,
Barcelona, Spain, September 2003.
2. Cayetano Guerra Artal. Contribuciones al seguimiento visual precategórico. PhD
thesis, Universidad de Las Palmas de Gran Canaria, Octubre 2002.
3. Hannes Kruppa, Modesto Castrillón Santana, and Bernt Schiele. Fast and robust
face finding via local context. In Joint IEEE Internacional Workshop on Visual
Surveillance and Performance Evaluation of Tracking and Surveillance (VS-PETS),
October 2003.
4. Henry Schneiderman and Takeo Kanade. A statistical method for 3d object de-
tection applied to faces and cars. In IEEE Conference on Computer Vision and
Pattern Recognition, 2000.
5. Moritz Storring, Hans J. Andersen, and Erik Granum. Physics-based modelling of
human skin colour under mixed illuminants. Robotics and Autonomous Systems,
2001.
6. Paul Viola and Michael J. Jones. Rapid object detection using a boosted cascade
of simple features. In Computer Vision and Pattern Recognition, 2001.
7. Ming-Hsuan Yang, David Kriegman, and Narendra Ahuja. Detecting faces in
images: A survey. Transactions on Pattern Analysis and Machine Intelligence,
24(1):34–58, 2002.
Evolutionary Color Constancy Algorithm based on the
Gamut Mapping Paradigm
Cristian Munteanu1,2, Agostinho Rosa1, Manuel Galan2, Enrique Rubio Royo2
1 Instituto de Sistemas e Robotica, Instituto Superior Tecnico, Av. Rovisco Pais 1, Torre
Norte 6.21, 1049-001, Lisboa, Portugal
{cmunteanu, acrosa}@isr.ist.utl.pt
http://laseeb.isr.ist.utl.pt/
2 CICEI, Universidad de Las Palmas de Gran Canaria, Edificio Central del Parque Científi-
Abstract. In recent years, extensive work has been done to design algorithms
that strive to mimic the robust human vision system which is able the perceive
the true colors and discount the illuminant from a scene viewed under light hav-
ing different spectral compositions (the feature is called “color constancy”). We
propose a straightforward approach to the color constancy problem by employ-
ing an Interactive Genetic Algorithm [1] (e.g. a Genetic Algorithm [2], [3]
guided by the user) that optimizes a well known and robust variant of color
constancy algorithm called “gamut mapping” [4]. Results obtained on a set of
test images and comparison to various color constancy algorithms, show that
our method achieves a good color constancy behavior with no additional
knowledge required besides the image that is to be color-corrected, and with
minimal assumptions about the scene captured in the image.
1 Introduction
The set of all possible responses due to known or expected surface reflectances, as
seen under a known, canonical illuminant (usually taken to be the daylight, or an
approximation of it), is a convex set, referred to as the canonical gamut. Similarly, the
set of responses due to an unknown illuminant is also a convex set. Illumination
change is assumed to be modeled by a simple diagonal transformation according to
the research of von Kries which was later confirmed to hold as a reasonable assump-
tion in [3]. It was shown that under the von Kries adaptation assumption the two
gamuts are within a diagonal transformation of each other. Due to the fact that GM
proposes a simple diagonal transformation to each pixel in the image, and because it
has been shown that under not very restrictive assumptions such diagonal transforma-
tions suffice for good color constancy [3], in this paper we adopt the GM method as
the basic color constancy transformation applied to correct a given image. The main
drawbacks of the classical GM methods are: a) they use extensive knowledge about
the camera used to take the picture; b) they use a priori information about the world
in which the image was taken in order to construct the gamuts.
How well a chromosome performs (that is, how good the diagonal mapping is in
recuperating the original colors and how well it corrects the input image) is judged
mainly by the human evaluator who looks at the corrected image and gives a numeric
score, called fitness value: f ( xi ) ranging from 0 (worst appearing image) to 10 (best
appearing image). The user shouldn’t evaluate all images corrected by each chromo-
some in the population, in each generation of the algorithm, because such a process
would become too tedious (there would be a lot of images to evaluate). Many of the
chromosomes to be evaluated are actually given a fitness value automatically, using a
clustering algorithm over fitness values previously allocated by the user.
t:=0
initialise P(t)
evaluate_1 P(t)
while (terminate(P(t)) ≠ true) do
{
P’(t) := select_tournament (P(t) | q)
P’’(t):= crossover(P’(t) | Pc)
P’’’(t):= mutate (P’’(t) | Pm)
P(t+1):= elitist(P’’’(t) ∪ P(t) | K)
evaluate_2 P(t+1)
t := t+1
}
endwhile
Fig. 1. Pseudocode of IGA-GM. t represents the time/generation index; P(t) is the population at
generation t, P’(t) is the population at generation t after selection, P’’(t) is the population at
generation t after selection and crossover, and P’’’(t) is the population at generation t after
selection, crossover and mutation. K is the number of elites.
3 Experimental results
The experimental part comprises two main sections: a) firstly we perform a com-
parison between IGA-GM and other Color Constancy algorithms for a single un-
known illuminant (e.g. different than canonical or daylight illuminant). b) secondly,
we check the effectiveness of IGA-GM on a set of scenes taken under two quite dif-
ferent than daylight illuminants.
Fig. 2. Comparison to Color Constancy algorithms (Detergent image). Upper-row from left to
right: original image (to be color-corrected), target image (image under canonical illuminant),
GW, Retinex; Lower-row from left to right: NN, GM, IGA-GM.
Table 2. RMS error calculated for Detergent image on pairs between target image (T) and
Color Constancy corrected image. The image that is not corrected (the original - input image)
is denoted as O.
Ball image
Fig.3 IGA-GM solutions on different illuminants. Figure is divided in horizontal bands for
each image/scene. For each band we have: first column first line: target image; second line first
column: image under “solux_4700+3202”; second line second column: IGA-GM solution;
second line third column: “IGA-GM solution + intensity correction”; third line first column:
image under “syl_wwf”; third line second column: IGA-GM solution; third line third column:
IGA-GM solution + intensity correction.
References
ykim@sejong.ac.kr
3 Hyundai Autonet Co. Ltd., Korea
Extended Abstract
During the last decade, the safety of occupants in the cars has been increased by im-
proving the airbag system, which can prevent life-threatening head (or chest) injuries
by avoiding direct contact to the dashboard during the accident. However, it is also
well known that the airbag deployment to the child (or infant) occupant could lead to
an injury [6]. To reduce the risk caused by airbag, it is important to design a smart
airbag that could control triggering and intensity of its deployment. In this paper, we
present a vision based automatic system, consisting of an occupant classification
system and an occupant’ pose recognition system as illustrated in Fig. 1 (a), by which
we aim to control whether the airbag should be triggered or not, and how strongly it
should be deployed when it is triggered.
(a) (b)
Fig. 1. The block diagram of the airbag control system (a) and the head tracking for the pose
recognition (b) - the solid circle indicates the head position of an occupant and the numbers
show the areas where the occupant can take a pose.
The occupant classification system can categorize the occupant into four differ-
ent classes such as the 5th%tile female (or adult), 6 year-old child, infant seat, and
empty seat as shown in Fig. 2. From a pair of stereo images acquired with a stereo-
camera, the disparity map is extracted by using a fast SAD (Sum of the Absolute
Difference) algorithm and a cubic model is used for background elimination [1]. And
the down-sampled image of the disparity map is used as an input for the SVM (Sup-
port Vector Machine) classifier [2]. Since the disparity maps of the 5th%tile female
and 6 year-old child are similar compared to other cases, the SVM classifier consists
of two stages: in the first stage, it classifies the occupant into three classes, such as
the 5th%tile female or 6 year-old child as the first class, the infant seat as the second
class, and the empty seat as the third class; in the second stage, the first class case is
classified as either the 5th%tile female or 6 year-old child. According to the classifica-
tion outcome, the system will deploy the airbag only if the occupant is classified as
the 5th%tile female, otherwise it will not be deployed.
For the pose recognition system, since we assume that the head of the occupant
can take the diverse poses, the passenger area is divided into 7 areas. And the system
is to recognize which area the occupant’ head belongs to, as shown in Fig. 1 (b). To
make the system robust, two algorithms are working in parallel as shown in Fig. 3.
First, we try to detect the occupant’ head in the given image by using a head contour
model [3] extracted from the accumulative difference images by combining a SVM
classifier that was trained using different sizes of head contours as illustrated in Fig. 4
(a) where each head contour model is made up different number of feature points, 97,
81 and 65, respectively. When the head detection succeeds, the system initializes
tracking parameters and repeats the sequence; otherwise, it checks whether the track-
ing parameters are initialized or not in the previous stage. For the head tracking, it
tracks the head by using an inverse compositional algorithm according to Lucas-
Kanade algorithm [4, 5]. When it fails, it counts as an error and returns to the head
detection stage. If the number of error becomes larger than the permitted number of
errors, the system halts and sends a warning signal.
Fig. 3. Flow chart of the pose recognition system
Fig. 4 (b) illustrates the detail schematic diagram of the head tracking. In the
initialization part, a model image is extracted from the given image and then it initial-
izes the tracking parameters, such as warping and motion parameters. In the head
tracking part, a warped image is made using the tracking parameters and it updates
the tracking parameters to minimize the sum of squared error between the model
image and the warped image. Such step is iterated until the tracking parameters con-
verge. Notice that the inverse compositional algorithm is fast enough to track the
head in the real-time base.
(a) (b)
Fig. 4. The head contour model with three different sizes of head (a) and the block diagram of
the head tracking algorithm (b).
The performance of occupant classification system was evaluated with our ste-
reo image database acquired within the experimental car. The stereo images totaled
802 and the resolution of each image was 320x240. We used down-sampled images
(resolution: 80x60) for reducing dimension of the data. Table 1 (a) shows average
results of occupant classification. The correction rate was 96.57% and 96.43% for the
case of RBF kernel and polynomial kernel, respectively. And, in case of classifying
the occupant at a time by using SVM classifier, the correction rate was 91.47%. Table
1 (b) shows results of occupant’ pose recognition. The head tracking experiment was
accomplished by using two image sequences which consist of 608 and 411 frames,
respectively. A success case was counted if any one of the two algorithms recognize
the occupant’ head as shown in Fig. 1 (b). Results show that the performance of the
system is satisfactory, suggesting that the vision based airbag control has a potential
in real applications. We plan to carry out further study to solve the occlusion problem
caused by thick coat, arm movement or hat.
References
1. H. G. Lee et al., Occupant Classification for Smart Airbag Using Stereovision and Support
Vector Machines, Springer Lecture Notes in Computer Science, Vol. 3173, pp. 642-647,
2004.
2. V. N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag New York, Inc.,
New York, NY, 1995.
3. M. W. Lee, Cohen I., S. K. Jung, Particle filter with analytical inference for human body
tracking, Motion and Video Computing, Workshop on., pp. 159-165, 2002.
4. B. Lucas, T. Kanade, An Iterative Image Registration Technique with an Application to
Stereovision, Proceedings of IJCAI 81, pp. 674-679, 1981.
5. S. Baker, I. Matthews, Lucas-Kanade 20 Years On: A Unifying Framework: Part 1. Techni-
cal Report, CMU-RI-TR-02-16, Carnegie Mellon University Robotics Institute, 2002
6. http://www.nhtsa.dot.gov
Wiener-type neuronal model with refractoriness
Extended Abstract
The extension of the Wiener neuronal model [3] in a way to include a con-
stant refractoriness time is the object of the present article. We shall denote by
{X(t), t ≥ 0} the Wiener process defined in I = (−∞, +∞) and characterized
by drift A1 (x) = µ and infinitesimal variance A2 (x) = σ 2 (µ > 0, σ > 0). To this
purpose, we assume that after the firing a period of refractoriness of constant
duration 1/ξ (ξ > 0) occurs, during which either the neuron is completely unable
to respond or it only partially responds to the incoming stimulations (cf., for in-
stance, [1], [2], [5]). We construct the return process {Z(t), t ≥ 0} in (−∞, S)
as follows (cf. [4]). Starting at a fixed state η ∈ (−∞, S) at time zero, a firing
takes place when X(t) attains for the first time a boundary S representing the
neuron’s threshold, after which the neuron is unable to fire again for the time
duration 1/ξ. At the end of the period of refractoriness, Z(t) is instantaneously
reset to η. The subsequent evolution of the process then goes on as described by
X(t), until the boundary S is again reached. A new firing then occurs, followed
by the period of refractoriness, and so on.
Hereafter, we shall provide a description of the random process {M (t), t ≥ 0}
representing the number of firings released by the neuron up to time t and
a characterization of the interspike pdf. To this purpose, for η ∈ (r1 , S), the
probability of having k firings up to time t is proved to be:
t−(k−1)/ξ
k−1
qk (t|η) := P {M (t) = k|Z(0) = η} = H t − g(k S, ϑ | k η) dϑ
ξ 0
t−k/ξ
k
−H t − g (k + 1) S, ϑ | (k + 1) η dϑ
ξ 0
Hence, the mean and variance of the interspike intervals are given by:
S−η
, k=0
µ σ 2 (S − η)
E(Ik ) = Var(Ik ) = , k = 0, 1, . . . .
1 S−η µ3
+ , k = 1, 2, . . .
ξ µ
Extension to random refractoriness in neuronal diffusion models will be the
object of future works.
References
1. Buonocore, A., Giorno, V., Nobile, A.G., Ricciardi, L.M.: A neuronal modeling
paradigm in the presence of refractoriness. BioSystems 67 (2002) 35-43
2. Esposito, G., Giorno, V., Nobile, A.G., Ricciardi, L.M., Valerio, C.: Interspike anal-
ysis for a single neuron’s activity in presence of refractoriness. In: Trappl, R. (ed.):
Cybernetics and Systems. Vol. 1. Austrian Society for Cybernetics Studies, Vienna
(2004) 199-204
3. Gerstein, G.L., Mandelbrot, B.: Random walk models for the spike activity of a
single neuron. Biophys. J. 4 (1964) 41-68
4. Ricciardi, L.M., Di Crescenzo, A., Giorno, V., Nobile, A.G.: An outline of theoretical
and algorithmic approaches to first passage time problems with applications to
biological modeling. Math. Japonica 50 No. 2 (1999) 247-322
5. Ricciardi, L.M., Esposito, G., Giorno, V., Valerio, C.: Modeling Neuronal Firing
in the Presence of Refractoriness. In: Mira, J., Alvarez, J.R. (eds): Computational
Methods in Neural Modeling. IWANN 2003. Lecture Notes in Computer Sciences,
Vol. 2686. Springer-Verlag (2003) 1-8
On Myosin II dynamics: from a pulsating
ratchet to a washboard potential?
Abstract
A Brownian motor for a Myosin II (“myosin”, in the sequel) head sliding
along an actin filament is designed. Various models based on the overdamped
equation s
U 0 [x(t), f(t)] 2kB T
ẋ(t) = − + Λ(t) (1)
ρ ρ
have often been proposed (see, for instance, [1], [2], [3]). Eq. (1) mimics the
myosin head as a Brownian particle subject to the impacts by the surrounding
molecules. In (1), Λ(t) is a standard Gaussian white noise, f(t) a random pro-
cess whose nature must be specified, and ρ the myosin friction drag coefficient.
U 0 [x(t), f(t)], the space derivative of a periodic potential U = U (x, t), stands
for the opposite of the global force arising from all deterministic interactions of
myosin head with the actin filament. The specification of f(t) marks the differ-
ence among the proposed models. In particular, in [1] the potential is assumed to
be an asymmetric saw-tooth function whose height switches between 2 different
values at Poisson distributed instants. In [2], U is a single saw-tooth potential,
the positions of whose minima change according to a Poisson law. Here, we shall
refer to this second formulation, by adding the assumptions that the motor step
equals the distance L between successive actin monomers (roughly 5.5 nm). Fur-
thermore, we assume that two 2L-periodic potentials, U1 (x) and U2 (x), alternate
according to a Poisson process, each potential being spatially shifted by L with
respect to the other. We denote by as the constant Poisson rate determining the
transition Us (x) → U3−s(x), s ∈ {1, 2}, of the particle’s position. The dynamics,
governed by s
Us0 (x) 2kB T
ẋ(t) = − + Λ(t) (2)
ρ ρ
is sketched in Fig. 1. Along the lines of [1], the transition pdf ps (x) and the corre-
sponding probability current Js (x) are determined within the reduced dynamics
in the stationary regime. It is shown that via certain averages with respect to
p(x) = p1 (x) + p2 (x), some features of the motor can be quantitatively deter-
mined. In particular, (i) the numerical solution of a fixed point equation (as
?
(Workshop “Computation and Simulation in Modelling Biological Systems”)
in [4]) leads to the stall force for the motor; (ii) by exploiting an idea of [5], the
derivative of the effective driving potential is calculated, and thus the intensity of
the force generated by the motor is determined; and (iii) the engine power with
reference to the reactions of 1 chemical cycle is calculated. Having determined
the force generated by the motor, via a Boltzmann equation the conditional
probability p of a forward step is estimated. Then, the mean duration µ between
pairs of successive steps is obtained in terms of p and J(x) = J1 (x)+J2 (x). Such
a value can be compared with that obtained via
Z 2L Z x
ρ V (x) V (y)
µ̂ = p dx exp dy exp − (3)
kB T 0 kB T x−2L kB T
providing the analogous quantity for a model based on washboard-type potential
(see [6]). The agreement between µ and µ̂ is pointed out by setting in (3) V (x) =
U (x) + Fe, where U (x) denotes the effective driving potential and Fe represents
a constant externally applied force.
3-s 3-s
Particles' displacement
. X
. X
L L
2L 2L
References
1. Astumian, R.D., Bier, M.: Fluctuation Driven Ratchets: Molecular Motors. Physical
Review Letters, Vol. 72, (1994) 1766–1769
2. Chauwin, J.F., Ajdari, A., Prost, J.: Force-Free Motion in Asymmetric Structures:
a Mechanism without Diffusive Steps. Europhysics Letters, Vol. 27, (1994) 421–426
3. Prost, J., Chauwin, J.F., Peliti, L., Ajdari, A.: Asymmetric pumping of particles.
Physical Review Letters, Vol. 72, (1994) 2652–2655
4. Chiu, C., Peskin, C.S.: Numerical analysis of the flux and force generated by a
correlation ratchet mechanism. Mathematical Biology, Vol. 44, (2002) 479–501
5. Wang, H., Oster, G.: Ratchets, power strokes, and molecular motors. Applied
Physics A, Vol. 75, (2002) 315–323
6. Lindner, B., Kostur, M., Schimanky-geier, L.: Optimal diffusive transport in a tilted
periodic potential. Fluctuation and Noise Letters, Vol. 1, (2001) R25–R39
Feedback effects in Stein’s coupled neurons
Extended Abstract
The dynamics of a pair of Stein’s neuronal units serially connected has been
recently analyzed ([4], [5]), under the assumption that they are subject to ex-
citatory and inhibitory stimuli with constant or with alternating rates. Making
use of a simulation procedure designed by some of us, various quantitative re-
sults have been obtained. In particular, it has been disclosed the existence of an
optimal value for the amplitude γ of the jumps of the membrane potential of the
second unit (the “receiving neuron”) attained as effect of the firing of the first
unit (the “sending neuron”). We also obtained results concerning the reaction
times, defined as the random times elapsing between a firing of the sending neu-
ron and the subsequent firing of the receiving neuron. The reaction times have
been analyzed in [5] under the assumption of constant inhibitory rate and for
both constant and alternating excitatory rate, with emphasis on the effects of
various choices of γ and of the amplitude of the refractory period.
The model considered in [4] and [5] is re-considered here under the new
assumption that a complete interaction is possible between the two neurons in
the sense that the firing of each neuron causes a jump of constant amplitude of
the membrane potential of the other neuron. The positive or negative value of
the jumps reflects the excitatory and inhibitory nature of firing neuron.
A first step of our study consists of the analysis of the interspike intervals
(ISI’s) of the two neurons by analyzing their histograms and empirical distri-
bution functions. Our results are then compared with those of [4], concerning a
serial connection between the two neurons.
To look at feedback effects in the firing activity of the two neurons we have
evaluated the relative entropy of the ISI’s distributions. Moreover, recalling the
statistical analysis performed by Elson et al. [6], we have focused our attention
on the reaction times and on ISI’s of the neurons in order to disclose their
synchronization. Some comparisons have been performed on the basis of the
values of the following index:
(2) (1) (1) (1)
Ri = (ti − tj )/(tj+1 − tj ),
(1) (2)
where tj is the firing time of the first neuron preceding the firing time ti of
the second one. Such an index measures the delay of the spike emission of the
second neuron ensuing the firing of the first neuron.
The synchronization of the firing activity of the neurons has been analyzed in
the special case in which one of the coupled neurons is excitatory and the other
is inhibitory. The relevance of the analysis of firing activity of pairs of neurons
under various mutual interactive paradigms is of interest since, for instance, it
is known that the cerebellar cortex includes pairs of coupled neurons, which has
motivated previous studies. (See, for instance, [2] and [3]).
Finally, we point out that our analysis of the synchronization of the firing
activity of the neurons has been performed also by studying the cross-correlation
function (see [1]). The auto-correlation function evaluated with reference to the
firing times of the two neurons has allowed us to disclose particular patterns of
spike trains from single neurons, including oscillating behaviors.
References
1. Borisyuk, R.: Oscillatory activity in the neural network of spiking elements. BioSys-
tems 67 (2002) 3–16
2. Burkitt, A.N.: Balanced neurons: analysis of leaky integrate-and-fire neurons with
reversal potential. Biological Cybernetics 85 (2001) 247–255
3. Cartling, B.: Control of computational dynamics of coupled integrate-and-fire neu-
rons. Biological Cybernetics 76 (1997) 383–395
4. Di Crescenzo, A., Martinucci, B., Pirozzi, E., Ricciardi L.M.: On the interaction
between two Stein’s neuronal units. In: Trappl, R. (ed.): Cybernetics and Systems
2004, Vol. 1. Austrian Society for Cybernetics Studies, Vienna (2004) 205–210
5. Di Crescenzo, A., Martinucci, B., Pirozzi E.: On the dynamics of a pair of coupled
neurons subject to alternating input rates. BioSystems (2004) (to appear)
6. Elson, R.C., Selverston, A.I., Huerta, R., Rulkov, N.F., Rabinovich, M.I., Abarbanel,
H.D.I.: Synchronous Behavior of Two Coupled Biological Neurons. Physical Review
Letters 81, N. 25 (1998) 5692–5695
Upcrossing first passage times for correlated
Gaussian processes
Extended Abstract
The first passage time (FPT) probability density function (pdf) of a stochas-
tic process through certain boundaries is used to model the first time when
the neuronal membrane potential crosses the threshold value thus generating
an action potential (see, for example, [1] and [5]). Along the line of previous
mathematical and computational approaches pursued to solve the so-called FPT
problem ([4],[2],[3]), here we investigate the asymptotic behavior of the FPT pdf
for a class of stationary Gaussian processes for increasing correlation times.
Let X(t) be a non-singular stationary Gaussian process with zero mean and
covariance E[X(t)X(τ )] = %(t − τ ) such that %(0) = 1, %̇(0) = 0 and %̈(0) < 0;
then, Ẋ(t), the derivative of X(t) with respect to t, exists in the mean-square
sense. Since X(t) is stationary, the correlation time is as ([5]):
Z +∞
τc := %(ϑ) dϑ.
0
We shall focus on the FPT upcrossing problem that we define as an FPT prob-
lem to a boundary S(t) ∈ C 1[0, +∞) for the subset of sample paths of X(t)
originating at time zero at a state X0 . The state is in turn viewed as a random
variable with pdf
f(x0 )
, x0 < S(0) − ε
P {X(0) < S(0) − ε}
γε (x0 ) ≡
0, x0 ≥ S(0) − ε,
where ε > 0 is a fixed real number and f(x0 ) denotes the normal pdf of X(0).
Then,
(ε)
TX0 := inf {t : X(t) > S(t)},
t≥0
is the ε-upcrossing FPT of X(t) through S(t) and its pdf is given by
Z S(0)−ε
∂ (ε)
gu(ε) (t) = P (TX0 < t) = g(t|x0) γε (x0 ) dx0 (t ≥ 0),
∂t −∞
where g(t|x0) denotes the FPT pdf of X(t) through S(t) conditional on X(0) =
x0 < S(0).
By using a straightforward variant of a method proposed by Ricciardi and
Sato ([6]) for the determination of the conditional FPT density, in ([2]) we have
obtained a series expansion for the upcrossing FPT pdf. However, the evaluation
of the partial sums is effectively impossible due to the outrageous complexity of
the involved functions and of their integrals. Nevertheless, if the correlation
function of X(t) satisfies
%̇(t) 1
lim %(t) = 1, lim %̇(t) = lim %̈(t) = 0, lim = − (1)
τc →+∞ τc →+∞ τc →+∞ τc →+∞ 1 − %2 (t) t
and if S(t) is a monotonically decreasing function, then the following approxi-
mate behavior holds as τc → +∞:
gu(ε) (t) ' −Ṡ(t) γε [S(t)], if S(t) < S(0) − ε, and gu(ε) (t) ' 0, otherwise.
An example of a process of concrete interest for applications that satisfies
(1) is characterized by the following covariance function:
1
%(t) = e−α|t|{cos(αω|t|) + sin(αω|t|)}, (t ∈ R) (2)
ω
where α > 0, ω ∈ R and τc = 2/ α(1 + ω2 ) .
In the present paper we indicate the practical usefulness of the above approx-
imation by means of reliable statistical estimations on numerically simulated
processes having covariance (2).
References
1. Di Nardo, E., Nobile, A.G., Pirozzi, E., Ricciardi, L.M.: On a non-Markov neuronal
model and its approximation. BioSystems, Vol. 48, (1998) 29–35
2. Di Nardo E., Nobile A.G., Pirozzi E., Ricciardi L.M.: Gaussian Processes and Neural
Modeling: an Asymptotic Analysis. In Cybernetics and Systems 2002, Vol. 1 (Trappl,
R., ed.), Proceedings of the 16th European Meeting on Cybernetics and Systems
Research, Vol I, Austrian Society for Cybernetics Studies, Vienna, ISBN 3-85206-
160-1, (2002) 313-318
3. Di Nardo E., Nobile A.G., Pirozzi E., Ricciardi L.M.: Computational Methods for
the Evaluation of Neuron’s Firing Densities. In Computer Aided Systems Theory
- EUROCAST 2003, (R. Moreno-Diaz, F. Pichler), Lecture Notes in Computer
Science 2809, Springer-Verlag Berlin Heidelberg New York, ISSN 0302-9743, ISBN
3-540-20221-8,(2003) 394-403
4. Giorno V., Nobile A.G., Pirozzi E., Ricciardi, L.M.: Non-stationary Gauss-Markov
Processes and Neuronal Modeling. In Cybernetics and Systems 2004 (Trappl R.,
ed.). EMCSR 2004 Austrian Society for Cybernetics Studies, Vienna, ISBN 3 85206
169 5, (2004) 211-215
5. Kostyukov, A. I., Ivanov,Yu.N., Kryzhanovsky, M.V.: Probability of Neuronal Spike
Initiation as a Curve-Crossing Problem for Gaussian Stochastic Processes. Biological
Cybernetics, Vol. 39, (1981) 157-163
6. Ricciardi, L.M., Sato, S.: On the evaluation of first–passage–time densities for Gaus-
sian processes. Signal Processing, Vol. 11, (1986) 339–357
Convergence of Iterations
Paul Cull
Extended Abstract
Convergence is a central problem in both computer science and in population
biology.
Will a program terminate?
Will a population go to an equilibrium?
In general these questions are quite difficult – even unsolvable.
In this talk we will concentrate on very simple iterations of the form
xt+1 = f (xt )
where each xt is simply a real number and f (x) is a reasonable real function.
Such one-dimensional iterations have special properties not shared by higher
dimensional iterations. For example, if f (x) is continuous and bounded, with
exactly one fixed point and no cycle of period 2, then the iteration will always
approach the fixed point.
For such a system, we say that an iteration is “globally stable” if it approaches
the fixed point for all starting points, and that the iteration is “locally stable” if
it approaches the fixed point for all initial points near enough to the fixed point.
Local stability can be checked by a simple derivative test, but showing global
stability requires more effort.
Population biologists have used one-dimensional iterations as simple models
of population growth. When they have shown local stability, they have assumed
global stability and have never run into any difficulty. Over the years a variety of
mathematical techniques have been used to show that specific simple population
models exhibit this
local stability implies global stability
phenomenon. But, no single method could prove this result for all the commonly
used population models.
We will show that there is a simple method which assures global stability.
Our method uses bounding of f (x) by a ratio of two linear functions. We call
this bounding “enveloping” and we show that
enveloping implies global stability.
Returning to the population models, we show that for these models local sta-
bility implies enveloping and hence global stability. For several models a single
enveloping function can be used for all parameter values, and we give a Tay-
lor series condition which implies this enveloping. For other models, different
enveloping functions are needed for different parameter combinations.
Although enveloping is a sufficient condition for convergence, we show by
example that enveloping is not necessary.
Finally we show that
enveloping implies global stability
even when f (x) is a discontinuous multi-function.
References
1. P. Cull. Global Stability of Population Models. Bulletin of Mathematical Biology,
43:47–58, 1981.
2. P. Cull. Local andGlobal Stability for Population Models. Biological Cybernetics,
54:141–149, 1986.
3. P. Cull. Local and Global Stability of Discrete One-dimensional Population Models.
In L. M. Ricciardi, editor, Biomathematics and Related Computational Problems,
pages 271–278. Kluwer, Dordrecht, 1988.
4. P. Cull. Stability of Discrete One-dimensional Population Models. Bulletin of
Mathematical Biology, 50(1):67–75, 1988.
5. P. Cull. Linear Fractionals - Simple Models with Chaotic-like Behavior. In D. M.
Dubois, editor, Computing Anticipatory Systems:CASYS 2001 - Fifth Interna-
tional Conference, Conference Proceedings 627, pages 170–181. American Institue
of Physics, Woodbury, N.Y., 2002.
6. P. Cull and J. Chaffee. Stability in discrete population models. In D. M. Dubois,
editor, Computing Anticipatory Systems: CASYS’99, pages 263–275. Conference
Proceedings 517, American Institute of Physics, Woodbury, NY, 2000.
7. P. Cull and J. Chaffee. Stability in simple population models. In R. Traappl, editor,
Cybernetics and Systems 2000, pages 289–294. Austrian Society for Cybernetics
Studies, 2000.
8. Paul Cull. Stability in one-dimensional models. Scientiae Mathematicae Japonicae,
58:349–357, 2003.
9. R. Devaney. An Introduction to Chaotic Dynamical Systems. Benjamin, Redwood
City, 1986.
10. M.E. Fisher, B.S. Goh, and T.L. Vincent. Some Stability Conditions for Discrete-
time Single Species Models. Bulletin of Mathematical Biology, 41:861–875, 1979.
11. M.P. Hassel. Density Dependence in Single Species Populations. Journal of Animal
Ecology, 44:283–296, 1974.
12. J.P. LaSalle. The Stability of Dynamical Systems. SIAM, Philadelphia, 1976.
13. T-Y. Li and J. Yorke. Period Three Implies Chaos. American Mathematical
Monthly, 82:985–992, 1975.
14. R.M. May. Simple Mathematical Models with Very Complicated Dynamics. Na-
ture, 261:459–467, 1976.
15. P.A.P. Moran. Some Remarks on Animal Population Dynamics. Biometrics, 6:250–
258, 1950.
16. A. Nobile, L.M. Ricciardi, and L. Sacerdote. On Gompertz Growth Model and
Related Difference Equations. Biological Cybernetics, 42:221–229, 1982.
17. C.J. Pennycuick, R.M. Compton, and L. Beckingham. A Computer Model for
Simulating theGrowth of a Population, or of Two Interacting Populations. Journal
of Theoretical Biology, 18:316–329, 1968.
18. W.E. Ricker. Stock and Recruitment. Journal of the Fisheries Research Board of
Canada, 11:559–623, 1954.
19. G. Rosenkranz. On Global Stability of Discrete Population Models. Mathematical
Biosciences, 64:227–231, 1983.
20. A. Sarkovskii. Coexistence of Cycles of a Continuous Map of a Line to Itself. Ukr.
Mat. Z., 16:61–71, 1964.
21. D. Singer. Stable Orbits and Bifurcation of Maps of the Interval. SIAM Journal
on Applied Mathematics, 35(2):260–267, Sept 1978.
22. S. Utida. Population Fluctuation, an Experimental and Theoretical Approach.
Cold Spring Harbor Symposium on Quantitative Biology, 22:139–151, 1957.
Semiautomatic Snake-Based Segmentation of
Solid Breast Nodules on Ultrasonography
Extended Abstract
From a seed point introduced by the specialist, the selection grows until a certain
threshold for the magnitude of the gradient is reached (see Fig. 1).
Afterwards, the contour of the selected area is considered as the initial snake
for the evolution of the active contours, based on the following equation:
∂u(x, y) 5u(x, y)
= g(I(x, y)) |5u(x, y)| div( ) + λ∇g(I(x, y)) 5 u(x, y) (1)
∂t |5u(x, y)|
1
g(I(x, y)) = q (2)
1 + α |∇Is (x, y)|2
References
1. Stavros, A.T., Thickman, D., Rapp, C.L., Dennis, M.A., Parker, S.H., Sisney, G.A.:
Solid breast nodules: use of sonography to distinguish between benign and malignant
lesions, Radiology 196 (1995) 123-134
2. Drukker, K., Giger, M.L., Horsch, K., Kupinski, M.A. and Vyborny, C.J.: Comput-
erized lesion detection on breast ultrasound. Medical Physics 29:7 (2002) 1438-1446
3. Revell, J., Mirmehdi, M. and McNally, D.: Applied review of ultrasound image fea-
ture extraction methods. In A Houston and R Zwiggelaar, editors, The 6th Medical
Image Understanding and Analysis Conference, BMVA Press (2002) 173—176
4. Kass, M., Witkin, A. and Terzopoulos, D.: Snakes: Active Contour Models. In 1st
International Conference on Computer Vision (1987) 259-268
5. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic Active Contours, International Jour-
nal of Computer Vision, 22:1 (1997) 61-79
6. Busse, L.J., Dietz, D.R., Glenn, W.M. jr: A non-linear algorithm for speckle reduc-
tion, IEEE Symposium on Ultrason. Ferroelec. and Freq. Contr., IEEE 91CH3079-1
(1991) 1105-1108
Parallel Progressive Multiple Sequence
Alignment
Erik Pitzer
References
1. Higgins, D.G., Sharp, P.M.: CLUSTAL: a package for performing multiple sequence
alignment on a microcomputer. Gene 73 (1988) 237–244
2. Thompson, J.D., Higgins, D.G., Gibson, T.J.: CLUSTAL W: improving the sensitiv-
ity of progressive multiple sequence alignment through sequence weighting, position-
specific gap penalties and weight matrix choice. Nucleic Acids Research 22 (1994)
4673–4680
3. Saitou, N., Nei, M.: The Neighbor-joining Method: A New Method for Recon-
structing Phylogenetic Trees. Journal of Molecular Biology and Evolution 4 (1987)
406–425
4. Needleman, S.B., Wunsch, C.D.: A General Method Applicable to the Search for
Similarities in the Amino Acid Sequence of Two Proteins. Journal of Molecular
Biology 48 (1970) 443–453
5. Myers, E.W., Miller, W.: Optimal alignments in linear space. Computer Applica-
tions in the Biosciences 4 (1988) 11–17
Computer Aided Simulation of Energy Limits
and Development of Bio-Thermodynamic Systems
Stanislaw Sieniutycz
Extended Abstract
In this research we focus on the theory and computer aided modeling of energy limits
in bio-thermodynamic systems with living organisms as the effect of extremum
properties of entropy and its associates. We also consider the role of physical laws in
establishing optimal forms and sizes of the organisms, and stress the universality of
these laws for animate and inanimate systems. For biological processes with variable
number of states an extremum principle provides quantitative picture of their
development. The results show a discrete gradient dynamics is governed by the
entropy potential. For an accepted complexity, some evolutions may progress in a
relatively undisturbed manner, whereas others may terminate in a rapid way due to
inherent instabilities, the features of variational principles for shortest paths and
transversality conditions. Numerical aspects of energy and size limits are described.
Recently, several approaches involving systems theory, computer science, optimal
control theory and thermodynamics have been applied, aimed at evaluation and
simulation of energy and size limits for bio-thermodynamic systems [1]. Static limits,
which refer to steady situations, were evaluated earlier; more complex limits
associated with dynamic processes were found later [1, 2]. Static limits may be
defined as maximum mechanical work produced or minimum work consumed by a
complex system or an organism at a steady state. Dynamic limits may be referred to
the same physical or biophysical quantities as static, yet they may include other terms,
e.g. those characterizing rates of the system’s evolution. Often, dynamic limits are
defined through extremum values of certain integrals, e.g. those describing action,
entropy production or availability dissipation. Usually dynamic limits are associated
with the system’s (organism’s) development, including phases of growth and aging.
Typical constraints at which the limits are operative are energy and mass balances.
Both limits are numbers referred either to steady-state points or to the whole evolution
of the system. Yet, static limits are functions of a steady-state point, whereas dynamic
limits are functionals of the strategy the system develops. In fact, the dynamic limits
are most often determined as extremum values of respective functionals.
In the course of its development, the system (organism) can change not only its
mass-energy outputs (inputs), but also the size, structure and shape. Size limits are
defined in terms of the extremum size that it can achieve in the course of its
individual development or during its evolution characterizing the progress through
generations. Growing systems (those that are capable of increasing their size and/or
number of states) can exhibit critical behavior when their size increases beyond a
certain value. In effect, some developments may progress in a relatively undisturbed
manner, whereas others may terminate in a rapid way due to instabilities. For majority
of nonlinear models their solutions may only be numerical; so we describe also basic
numerical aspects of calculations associated with the computer modeling of the limits.
The structure of the research is as follows. In its basic part energy and size limits
are treated. In particular, limits for steady power yield or consumption are considered
and a problem of dynamical energy limits is formulated. This part terminates with the
scheme of multistage production of mechanical energy (work) in a sequence of
animate or inanimate energy generators. Next we refer to a group of works [1, 3] that
analyze the evolution of living organisms treated as multistage systems by the
complexity criterion based on a potential, this potential being the entropy itself or an
entropy-based complexity. Classical thermodynamic quantities do not appear, yet the
model yields an extremum principle that, as in thermodynamics, implies an extremum
of a potential. Dual forms of this principle are also considered. Next we develop a
discrete, nonlinear generalization of dynamics in metric spaces that may be
curvilinear. Numerical approaches applying the dynamic optimization methods and
leading to unsteady-state limits are then described and, in particular, sequential
systems with growing number of stages are treated by methods of dynamic
programming and maximum principle. Suitable dynamic programming algorithms
(Bellman’s recurrence equations) are derived and computer-aided simulations of their
solutions are discussed. In the final part we conclude that many features of living
organisms can be predicted by describing their complex evolutions in terms of
variational principles for shortest paths along with suitable transversality conditions.
Energy, size and shape limits. Each organism can be regarded as a complex system
with very many elementary objects, links and couplings. Some inputs serve to supply
the consumed energy and substances to the organism (system) whereas some outputs
(e.g. muscles) are able to deliver produced mechanical energy out of the organism. In
fact, there is an analogy between an energy system of inanimate Nature and a living
organism. There are also at least two essential differences between the living system
and a machine: the first one is the much larger complexity of living system than a
machine and the second is the presence of huge number of information streams out
and within the living system. Yet, the essential property of living systems, the
capability to work, is typical of work generators. By applying physical principles we
are nowadays able to evaluate energy and dimension limits for living organisms and
describe behavior of many organs. While living systems have developed various
strategies to manipulate their organization in order to satisfy the principles of both
physics and biology, it is the physical laws that set limits to their size, functioning and
rate of development. For example, the physical law of thermal conduction sets the
size of warm-blooded aquatic animals that require a minimum diameter in order to
survive in oceans. Land animals that are able to run have body masses ranging from
little mice of M= 10-2 kg to big ungulates weighing about half ton (0.5x103 kg).
Aquatic mammals span the range from M= 101 kg baby seals to 105 kg blue whales.
Flying birds size from 10-2 kg humming birds to 15 kg mute swans [4]. These results
are adduced following Ahlborn [4] who has shown that the size limitations are
consequence of thermodynamics which sets the energy budget and power production
of animals. The power of all animals in terms of their body mass can be described by
a unique metabolic law. Species in ecosystems funnel energy into their own
production or reproduction and influence catalytic processes in the ecosystem.
A number of recent developments in thermodynamic optimization focus on the
generation of optimal geometric form (shape, structure, topology) and shape limits in
animate and inanimate flow systems, where the flow configurations are free to vary.
The principle that generates geometric form searches for best global performance (e.g.
minimum irreversibility) subject to global finiteness constraints (volume, weight,
time). The resulting structures obtained in this way are named constructal designs [5].
The same objective and constraints principle applies for the optimally shaped flow
paths that occur in natural systems (animate and inanimate) which is Bejan’s
constructal theory [5]. Examples are obtained from engineering: the distribution of
heat transfer area in power plants, optimal sizing and shaping of flow channels and
fins, optimized ratios of heat exchanger structures, aerodynamic and hydrodynamic
shapes, tree-shaped assemblies of convective fins, tree-shaped networks for fluid
flows, optimal configurations for streams with bifurcation, insulated pipe networks
distributing hot water and exergy over a fixed area, and distribution networks for
entities transferred in societies (goods, currency, information). The principle-based
generation of flow geometry unites optimization developments known in mechanical
engineering with those in civil engineering and social organization. These results,
obtained in A. Bejan‘s research group [5], generalize methods of thermodynamics, as
they show how their principles can mimic the development of optimal configurations
in engineering and social organization. The organization is understood in a broader
sense, in particular as the one that refers to populations of living organisms (e.g. those
in ecosystems). The results show that flow systems with geometric structure exhibit at
least two flow regimes, one with high resistivity (slow, diffusion, walking) that fills
the volumes of smallest finite scale, and one or more with low resistivity (e.g., fast,
streams, channels, streets). The optimal balancing of the regions with different flow
regimes means that the material and channels must be distributed in certain ways.
Better global performance is achieved when the distribution is relatively uniform, in
spite of the large differences between the high- and the low-resistively domains [5].
Development and evolution of species. It is the difference between the maximum
entropy and the actual (observed) entropy that quantifies the organization in a
complex system. This property refers to the various forms of organization: with
respect to characters over species, cells over all cell types, organisms over genotypes.
Moreover, it is the entropy-based complexity, rather than fitness or organization, that
determines direction of evolutionary processes. Spontaneous increase of complexity
is regarded as a basic feature of evolution. For biological systems a reasonable
measure of complexity is the different number of components they contain, so as to
be consistent with well-known Williston's law which predicts that very similar
components will either merge or specialize. Staying on the ground of the idea of
increasing complexity, the system's organization acts as a force which prevents loss
of components and allows complexity increase by the integration of new ones. When
speaking about the complexity and related entropy one issue is particularly important:
a nonequililibrium entropy has to be necessarily applied because the difference
between the maximum entropy and the actual (observed) entropy governs the
organization in a complex system. Here we consider the idea of the extremum
development of complexity or complexity-related criterion of evolution as applied to
some recent information-theoretic models for multistage living systems characterized
by sequentially evolving states. For these models the evolution of living organisms is
a multistage effect governed by the complexity criterion Γ based on information-
theoretic entropy. An original optimization problem of maximum change of entropy
or entropy-related complexity Γ between a point (for a system with n states) and the
surface of constant distance from the point can be formulated and solved by
variational calculus. For practical purposes also a dual of this problem is considered
which is the problem of minimum distance between a point (for a system with n
states) and the entropy manifold of the system with n+1 states. In this formulation, an
s-dimensional evolution can be obtained via minimization of the length functional
where the independent probabilities pi are constrained to reside on constant-
complexity manifold satisfying the constraint dΓ=0. It may then be shown that both
the original problem of maximum complexity and its dual problem of the minimum
distance yield a discrete evolution dynamics satisfying the Onsager-like form
∂Γn+1( p0,. ps) ∂Γn+1( p0,. ps) ∂Γn+1( p0,. ps)
ps (n +1) − ps (n) = Ls1 + Ls2 ...+ Lss , (1)
∂ p0 ∂ p1 ∂ ps
where Lik = gik∆τ. The symmetry matrix Lik is the well-established result of
irreversible thermodynamics. Therefore the evolution process can be imbedded quite
naturally into a large family of thermodynamic processes.
Conclusions. We defined energy, size and shape limits and extremum properties of
complex systems (living organisms). We described numerical aspects of calculations
associated with the computer aided modeling of these limits. We have shown that the
development of living organisms can be described in terms of variational extremum
principles along with suitable transversality conditions. We found gradient dynamics
that governs the evolution problems, the result that is consistent with the entropy
principle of extremality as the driving factor in the discrete dynamics of complex and
living systems, postulated recently [3]. We conclude that the extremum principle
should be integrated with the theory of behavior of biochemical systems.
Acknowledgments. The author acknowledges the financial support from the KBN,
grant T09C 024 26: Non-equilibrium thermodynamics and optimization of chemical
reactions in physical and biological systems.
References
1. Berry, R.S., Kazakov V.A., Sieniutycz S., Szwast Z., Tsirlin A. M.: Thermodynamic
Optimization of Finite Time Processes, Wiley, Chichester (2000).
2. Jorgensen, S.E., (ed.): Thermodynamics and Ecological Modelling, Lewis Publishers of
CRC Press, Boca Raton (2001).
3. Gontar, V.: Entropy principle of extremality as a driving force in the discrete dynamics of
complex systems, Chaos, Solitons and Fractals 11 (2000) 231-236.
4. Ahlborn B.K., Thermodynamic limits of body dimension of warm-blooded animals, J. of
Non-Equilib. Thermodyn. 40 (1999) 407-504.
5. Bejan, A., Shape and Structure from Engineering to Nature, Cambridge University Press,
Cambridge (2000).
Soft Computing and Geometrical Control for
Computer Aided Driving
Abstract. After having completed the design of control systems for real
autonomous cars in urban environment using straight lines as reference, we are
now intending to build a fuzzy control system based in clothoids [2], the curve
used in roads and train ways. This work proposes a method based on soft
computing and upgraded via genetic algorithms. Both sets of simulations are
shown, and compared.
Introduction
One of the big efforts in a control system is to reduce the force applied on the
actuators. In autonomous vehicles the amount of electrical required is an important
issue. Our goal is to reduce the electrical expense in the direction actuators. Moreover
the reduction of excessive steering wheel movements increases the comfort feeling.
We have proposed and tested a control method for vehicle navigation in curved
sections. We have used a curvature control method based on fuzzy logic with genetic
algorithms optimization over a geometric trajectory design. We have obtained
impressive results on simulation and now we are testing these results on real
autonomous vehicles.
Nowadays most roads are designed trying to maintain a progressive and smooth
trajectory in its curved sections. This designing method is explained in “Carreteros”
web page [4]. Each curve is divided into three parts, in the first one the curvature of
the road increases, in the second one the curvature is maintained constant and in the
last part the curvature is reduced to cero. Based on this idea we design a trajectory for
each curve. Following this trajectory the vehicle smoothly changes its curvature from
cero to a given number, and then from this number to cero.
To be more precise in [3] this differential equation system is shown relating the state
and control variables and its integration equations for admissible paths:
(x ′ y ′ θ ′ κ ′) = v ⋅ (cos θ sin θ κ 0 ) + σ ⋅ (0 0 0 1) (1)
where (x y θ κ) stand for attitude and curvature, and σ is σ= κ’= Φ/cos2Φ. (Φ is the
wheel angle).
Our proposal is to express position of the car depending on its curvature,
f(x,y,κ,v,t) related to steering angle and control the car using its usual inputs,
acceleration and curvature (v’, κ=1/r).
The membership function parameters are improved with a genetic algorithm fitness
function that combines the accumulated effort and the maximum error made on the
circuit curves. We have obtained different optimized set values depending on the
nature of the circuit curves (motorways with small curvature or race circuits).
The simulator uses detailed maps of competition circuits, mainly Xerez, with its
real proportions. And it takes care of our cars, CITROËN, mechanical characteristics.
We have carried out several simulations in this circuit, and the results show a drastic
effort reduction after the genetic algorithm optimization.
Left fig. Differences between two consecutive steering wheel actuator goals. The blue one is
the steering wheel actuation without optimization and the pink one is after application of
genetic algorithm. Right fig. Drawing a course on Xerez circuit map
References
Jesús Nuevo1, Miguel Ángel Sotelo1, Manuel Ocaña1, Luis Miguel Bergasa1
1Department of Electronics, University of Alcalá, Alcalá, Madrid, Spain
{jnuevo, michael, mocana, bergasa}@depeca.uah.es
1 Introduction
2 System Description
Lane Tracking
The system is divided in three modular subsystems with specific functions. The
first subsystem is responsible for lane detection and tracking, as well as lane crossing
monitoring. Images obtained from the camera are processed and clothoid curves are
fitted to the detected markings. The algorithm scans up to 25 lines in the area of in-
terest, from 2 meters in front of the camera position to below the horizon. The devel-
oped algorithm implements a non-uniform spacing search that reduces certain unsta-
bilities in the fitted curve. The final state vector is composed of 6 variables [7] for
each line on the road: coh, c1h, cov, c1v, xo, Ρo, where coh and c1h represent the clothoid
horizontal curvature parameters, cov and c1v stand for the clothoid vertical curvature
parameters, while xo and Ρo are the lateral error and orientation error, respectively,
with regard to the centre of the lane. The clothoid curves are then estimated based on
lane marking measurements using a Kalman filter for each line. These lines conform
the area of interest. Figure 1 depicts a sequence of images in which the result of the
lane tracking algorithm is overprinted on the road images.
Fig. 1. Lane tracking example in a sequence of images. The green lines represent the estimated
lines of the road. The example also depicts the error between the left wheel of the car the the
left lane (left), the error between the right wheel of the car and the right lane (right), the radious
of curvature of the road estimated at a lookahead distance of 50m (R), and the maximum rec-
ommended velocity to bend the curve (V) according to the radious of curvature.
The area of the image contained between the lines that represent the lane is
scanned in order to look for candidate vehicles along the lane as depicted in figure 2.
Car detection is performed in the areas limited by lane markings. Based on character-
ics of shape, size and position, among others, candidates are selected. Each one of
the candidates is then divided into five different subareas, that are then processed
independently. For each subarea of the candidate, a support vector machine (SVM)
[6] has been trained, using the tool ‘TsetBuilder’ we have developed. The areas are
validated independently, and another SVM determines whether the outputs of the first
SVMs represent those of a valid cadidate. Candidates are tracked using a Kalman
filter. Figure 3 shows the result of the detection and tracking algoritm.
The distance between the ego-vehicle and the preceding vehicle along the lane be-
comes the input to the Adaptive Cruise Control System.
Acknowledgements
References
Extended Abstract
Optimization is one of the most common and pervasive issues in real-world engi-
neering and economic systems and it is at the heart of any decision-making task
in which a choice must be made between several alternatives to achieve multi-
ple, sometimes conflicting, objectives. The objectives are generally formalized as
analytical functions or performance indices.
The technical literature on optimization methods is really extensive, as this
fundamental subject has received a tremendous amount of attention since it
came of age about 50 years ago. Single-objective optimization is by far the most
researched problem in this field, although many real-life situations are multi-
objective optimization problems per se. A multi-objective optimization problem
is very often converted into a single-objective optimization case by integrating
the multiple performance indices into a single one [1].
The standard solution for truly multi-objective optimization problems is to
find the so-called Pareto-optimal front. The Pareto front is formed by the solu-
tions in which any change in any of the decision variables aimed at improving a
particular performance index will produce a deterioration in some of the other
performance indices. Due to the inherent difficulties in calculating the analytical
Pareto optimal surfaces for many real-world systems, evolutionary methods have
lately been applied to solve multi-objective optimization problems [2, 3].
In most of the multi-objective optimization problems that have been tackled
by evolutionary techniques so far, the performance indices to be optimized:
Ji (x) = Ji (x1 , x2 , . . . , xn )
(1)
x ∈ Rn ; i = 1, 2, . . . N
are not time-varying and the decision varibles x1 , x2 , . . . xn have static con-
straints:
gi (x) ≥ 0 ; i = 1, 2, . . . p
(2)
hi (x) = 0 ; i = 1, 2, . . . q
Only very recently have a number of papers dealing with the application
of evolutionary computation to solve multi-objective optimization cases been
published, in which either the performance indices Ji (x), the decision variables
x or both are time dependent and dynamic [4, 5]. Most of this research work
refers to variations on static optimization problems, in which some well-known
static benchmark optimization functions are exposed to time-varying dynamics.
Noise in the measurement of the performance indices or in the decision variables
is sometimes also added.
In this paper, we present an engineering problem (automatic car parking)
that ultimately leads to the formalization of an active multi-objective dynamic
optimization problem as a cooperative game. However, instead of using evolu-
tionary controllers to search for an optimal, or at least efficient, set of if-then
reasoning rules, we propose a much simpler method based on what we call a
biomimetic approach to sensory-motor coordination. Sensory-motor coordina-
tion is a central issue in the design of most real-world engineering systems, as
the designer somehow has to deal with the coordination of perceptual informa-
tion and physical actions. We have already applied this biomimetic approach
to solve specific manipulation and locomotion problems in articulated mecha-
nisms [6, 7] and wheeled vehicles [8, 9]. In this paper we illustrate our approach
to sensory-motor coordination with the problem of automatic parallel parking
for autonomous vehicles.
References
1. Deb, K. (2001) Multi-Objective Optimization Using Evolutionary Algorithms. Wi-
ley, New York
2. Fonseca, C.M., Fleming, P.J. (1993) Genetic algorithms for multiobjective opti-
mization: Formulation, discussion and generalization. Proc. of the 5th Int. Conf. on
Genetic Algorithms, 416–423
3. Coello, C.A. (2003) Special Issue on Evolutionary Multiobjective Optimization.
IEEE Trans. on Evolutionary Computation, 7:2, 97–99
4. Branke, J. (2002) Evolutionary Optimization in Dynamic Environments, Kluwer,
Boston
5. Jin, Y., Sendhoff, B. (2003) Connectedness, regularity and the success of local search
in evolutionary multi-objective optimization. Proc. IEEE Congress on Evolutionary
Computation (CEC-2003), 1910–1917
6. Maravall, D., De Lope, J. (2003) A bio-inspired robotic mechanism for au-
tonomous locomotion in unconventional environments. In C. Zhou, D. Maravall, D.
Ruan (Eds.) Autonomous Robotic Systems: Soft Computing and Hard Computing
Methodologies and Applications. Physica-Verlag, Springer, Heidelberg, 263–292
7. De Lope, J., Maravall, D. (2004) A biomimetic approach to the stability of biped
robots. Climbing and Walking Robots. Springer-Verlag, Heidelberg
8. De Lope, J., Maravall, D. (2003) Integration of reactive utilitarian navigation and
topological modeling. In C. Zhou, D. Maravall, D. Ruan (Eds.) Autonomous Robotic
Systems: Soft Computing and Hard Computing Methodologies and Applications.
Physica-Verlag, Springer, Heidelberg, 103–139
9. Maravall, D., De Lope, J., Patricio, M.A. (2004) Competitive Goal Coordination
in Automatic Parking. Proc. 1st of the European Workshop on Evolutionary Algo-
rithms in Stochastic and Dynamic Environments (EvoSTOC-2004)
Intelligent Automation of Electric Power Steering
Systems
1 Introduction
References
1. R. García et al., “Frontal and Lateral Control for Unmanned Vehicles in Urban
Tracks”, IEEE Intelligent Vehicles Symposium, Versailles, France, 2002.
2. JE. Naranjo et. al., “Adaptive Fuzzy Control for Inter.-Vehicle Gap Keeping”,
IEEE Trans. on ITS,Volume 4: No. 3, September 2003, pp. 132-142.
3. J.E. Naranjo, et al., “A Throttle & Brake Fuzzy Controller: Towards the Automatic
Car”, LNCS 2809, Springer-Verlag, Julio 2003, pp 291-301.
4. A: Broggi et al., “The ARGO Autonomous Vehicles Vision and Control Systems”,
International Journal of ICS, Vol. 3, No. 4, pp. 409-441, 1999.
5. J.E. Naranjo et al., “Overtaking Maneuver Experiments with Autonomous Vehi-
cles”, Proc. 11th ICAR 2003, Coimbra, Portugal, pp. 1699-1703, June 2003.
6. T. Endo et al., “Development of reverse parking assist with automatic steering”,
ITS 2003 Congress Proceedings, Madrid, November 2003.
Development of an Artificial-Vision Controlled Traffic-
Light Prototype
1 System Description
The system consists of two cameras situated at only one signal post placed at an inter-
section. One of them focuses on the pedestrian crossing while the other focuses on
the vehicles arriving grid, as seen in Figure 1. The prototype has been developed as to
be carried out by a conventional PC. A TWIDO programmable automaton, from the
Schneider Electric firm, is in charge of the traffic light control.
Fig. 1. General scheme of the prototype. Two cameras: one of them focussing on the pedestri-
ans and the other one on the vehicles. Images are then sent to the computer and as result a final
working mode is sent to the automaton responsible for the traffic light switch.
Two different types of algorithms have been developed. Low level algorithms are
in charge of obtaining the moving objects and assigning them into connected compo-
nents. High level algorithms are in charge of providing the system with crucial infor-
mation like position and trajectory, to make the required decision.
Kalman filter is used for tracking moving objects, both vehicles and pedestrians.
An extended Kalman filter has been implemented [2, 5]. The algorithm adapts its
model at each step to improve the movement estimation of the component. Position,
velocity and values of the acceleration at different steps are used to predict the future
object location. Gathering the information provided by the trajectories and different
crucial zones established such as the crossing area and the vehicles grid, a decision
can be made on the operating mode to be sent to the automaton. These crucial zones
are fixed by initial parameters that have to be set in accordance with the particular
geometry of the crossing area.
System works well under real conditions, without influencing the presence of
cameras in the behaviour of pedestrians and vehicles. After the multiple tests done
under RACE supervision, we can fairly state that the artificial techniques used in this
prototype are reliable and dependable to be introduced in our current daily life in
roads and pedestrians crossings. The enhancement of the methods and algorithms
used, allowing the decisions to be made in real time, make the system mature and
trustable enough.
Behaviour of vehicles in the grid can be analysed, and an option to control density
of traffic in different ways has been considered.
References
Extended Abstract
There are several strategies of how to retrieve depth information from a sequence of
images, like depth from motion, depth from shading and depth from stereopsis.
Brown et al. [1] describe a good approximation to all of them in their survey article.
Koenderink and van Doorn [2] expressed the necessary theory in their initial works
related to disparity restriction, and Wildes [3] implemented some of their ideas [4].
More recently, disparity in stereoscopy continues showing its great interest (e.g., [5]).
In this paper, as a novelty, we have chosen not to use direct information from the
image, but rather the one derived from motion analysis. The system proposed uses as
input the motion information of the objects present in the stereo scene, and uses this
information to perform a depth analysis of the scene.
The input to our system is a pair of stereo image sequences. The sequences have
been acquired by means of two cameras arranged in a parallel configuration. The
central idea behind this approach is to transpose the spatially defined problem of
disparity estimation into the temporal domain and to compute the disparity simultane-
ously with the incoming data flow. This can be achieved realizing that in a well-
calibrated fronto-parallel camera arrangement the epipolar lines are horizontal and
thereby identical to the camera scan-lines. Our team has already tested the motion
analysis algorithm used in this work in monocular video sequences ([6]-[8]).
In this case, motion analysis performs separately on both stereovision sequences in
two phases. The first analysis phase is based in grouping neighboring pixels that have
similar grey levels in closed and connected regions in an image frame. The method
used is segmentation in grey level bands. This method consists in reducing the resolu-
tion of illumination levels of the image, obtaining this way a lower number of image
regions, which potentially belong to a single object in motion. The second phase has
to detect possible motions of the segmented regions through the variation of the grey
level band of the pixels.
After motion detection, we now introduce a representation that may help to estab-
lish further correspondences between different motion information. This representa-
tion finds its basis in the permanency effect. This effect considers the jumps of pixels
between bands, and it consists of a matrix of charge accumulators. The matrix is
composed of as many units in horizontal and vertical direction as pixels there are in
an image frame. Initially all accumulators are empty; that is to say, their charge is
zero. When a jump between grey level bands occurs at a pixel, the charge unit (accu-
mulator) of the permanency memory at the pixel’s position is completely charged.
After the complete charge, each unit of the permanency memory goes decrementing
with time (in a frame-by-frame basis) down to reaching the minimum charge value,
while no motion is detected, or it is completely recharged, if motion is detected again.
Figure 1 shows all these issues. Figure 1a and 1b show two images of a monocular
sequence. The advance of a car may be appreciated, as well as a more slight move-
ment of a pedestrian. In figure 1c you may observe the effect of these moving objects
on the permanence memory.
Fig. 1. Permanency effect: (a) one image of a sequence, (b) same perspective after some sec-
onds, (c) motion trails as represented on the permanence memory.
The difference between a quick object as the car, which is leaving a very long mo-
tion trail (from dark grey to white), and a pedestrian whose velocity is clearly slower
and whose motion trail is nearly unappreciable with respect to the cars one, is pre-
sented. Thus, permanency memories enable representing the motion history of the
frames that form the image sequence, that is to say, there is segmentation from the
motion of the objects present in the scene.
Now, motion-based segmentation facilitates the correspondence analysis. Indeed,
motion trails obtained through the permanency memories charge units are used to
analyze the disparity between the objects in the stereo pair. The set of all disparities
between two images of a stereo pair is denominated the disparity map. The key idea is
that a moving object causes two identical trails to appear in epipolar lines of the per-
manency stereo memories. The only difference relies in their relative positions, af-
fected by the disparity of the object at each moment.
Looking at figure 2 it is possible to analyze the motion of each one of the three ob-
jects present in the permanency memories from their motion trails. This initial analy-
sis is independent of the epipolar constraint studied. You may observe that object “a”,
which has a long trail and has his maximum charge towards the left, is advancing to
the left at a high speed. Object “b”, with a shorter trail, is also advancing towards the
same direction but at a slower velocity. Finally, object “c”, whose trail is inverted in
horizontal, is moving to the right at a medium velocity, as shown by its trail.
Also from figure 2, but now comparing between the motion trails in both epipolar
lines, disparity is analyzed. Motion trail of object “b” presents a null disparity. There-
fore, we can conclude that this trail corresponds to an object that is far away from the
cameras. Remember that due to our parallel cameras configuration, pixels with a null
disparity are located in the infinite. Object “a” has a little greater disparity. Finally,
object “c“ offers the greatest disparity.
Left perma-
nency memory
Right perma-
nency memory
a b c
Acknowledgements
References
1. Brown, M. Z., Burschka, D. & Hager, G. D., “Advances in Computational Stereo”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 8, (2003).
2. Koenderink, J.A. & van Doorn, A.J., “Geometry of binocular vision and a model for stere-
opsis”, Biological Cibernetics, vol. 21, (1976): 29-35.
3. Wildes, R. P., "Direct recovery of three-dimensional scene geometry from binocular stereo
disparity", IEEE Transactions on Pattern Analisis and Machine Intelligence, vol. 13, no. 8,
(1991): 761-774.
4. Wilson, R. & Knutsson, H., “A multiresolution stereopsis algorithm based on Gabor repre-
sentation”, Proceedings of the IEE International Conference on Image Processing and Ap-
plications, (1989): 19-22.
5. Gutiérrez, S. & Marroquín, J.L., “Robust approach for disparity estimation in stereo vision”,
Image and Vision Computing, vol. 22, no. 3, (2004): 183-195.
6. Fernández-Caballero, A., Fernández, M.A., Mira, J., Delgado, A.E., “Spatio-temporal shape
building from image sequences using lateral interaction in accumulative computation”, Pat-
tern Recognition, vol. 36, no. 5, (2003): 1131-1142.
7. Fernández, M.A., Fernández-Caballero, A., López, M.T., Mira, J., “Length-speed ratio
(LSR) as a characteristic for moving elements real-time classification”, Real-Time Imaging,
vol. 9, (2003): 49-59.
8. Fernández-Caballero, A., Mira, J., Delgado, A.E., Fernández, M.A., “Lateral interaction in
accumulative computation: A model for motion detection”, Neurocomputing, vol. 50,
(2003): 341-364.
Pedestrian Detection for Intelligent Vehicles based on
Active Contour Models and Stereo Vision
Intelligent Systems Lab. Universidad Carlos III de Madrid. Leganes, Madrid, 28911. Spain
{chilario, jcollado, armingol, escalera}@ing.uc3m.es
Extended Abstract
In last years, the European Union, Member States and the automotive industry have concen-
trated their efforts on improving road safety [1][2]. Avoiding accidents due to human errors is
one of the main objectives of Advanced Driver Assistance Systems (ADAS). European traffic
accidents figures show a very large number of collisions between pedestrians and vehicles.
Every year, more than 200,000 pedestrians are injured, with over 6,000 killed. In spite of the
incredibly high number of victims, the protection of pedestrians has received little attention
by the research community [3]. The projects that have dealt with this problem are quite re-
cent, as has been pointed out at the Fifth Framework Programme [1].
The algorithm proposed in this paper, relies on the information provided by computer vi-
sion for the development of a pedestrian protection system [4]. The use of visual information
is a promising approach to cope with the different appearances of pedestrians and changes of
illumination in a cluttered environment. Although other sensors as lasers or radars can detect
pedestrians, vision is the unique that can comprehend their motion and predict their move-
ments. An active contour model is used to segment pedestrians from images taken by an on-
board camera. A stereo module is integrated to guide the location of active contours as they
are very sensitive to their initial position. Stereo vision is used to generate a disparity map of
the scene. In order to filter the disparity map, distance measures are employed and vertical
symmetries are looked for in the filtered area, avoiding a complete scan of the image. Those
regions with a high vertical symmetry are potential candidates for an active contour initializa-
tion.
Fig. 1. [Left]A detail of the Stereo Vision System used for the stereo module. [Right] On-
board PCs for the visual information processing.
Fig. 2. [Top-left] A right image taken by the Stereo System. [Top-right] Disparity map, [Bot-
tom-left] vertical symmetries and [Top-right] vertical edges of that image.
Active contours or “snakes” were proposed by Kass et al [5]. Its ability to extract contours
even in presence of gaps or occlusions, jointed to its dynamic behavior, makes this approach
adequate for non-rigid objects detection and tracking. The deformable model developed ex-
tends the greedy algorithm of Williams and Shah [6]. Their model maintains the points in the
snake more evenly spaced than Kass [5]. Moreover, the discretization of the model proposed
is controlled by inserting or removing points whenever the distance between adjacent points
became too large or too close. This way, the natural tendency to shrink of the snakes is com-
pensated. Several measures have been employed to avoid shrinkage, as controlling the
amount of stretching and bending of the contour locally or modifying the internal energy
formulation.
On the other hand, a new internal force is included to control the shape of the deformable
model. This regularizing force prevents the shrinking effect of the snake, as it is based on
high degrees of smoothness of the contour. It looks for convex and concave regions in the im-
age which are prone to correspond to the head and feet of a pedestrian. For the external
forces, a new potential field which smoothly extends over a long distance is defined. In order
to mitigate the influence of other contours with the same features as the ones rewarded by the
model, more specific knowledge about the target object is added. As pedestrians have a
strong vertical symmetry, the algorithm looks for points in the image that being symmetric,
are near enough from its vertical axis. Demanding furthermore that points being symmetric
and near a vertical axis must belong to a vertical edge, results are improved. Besides this po-
tential field, the image gradient and distances to vertical borders are also considered.
Once the snake is initialized on an object contour in the first frame, it will automatically
track the contour from frame to frame. This approach requires that the deformation and
movement of an object between frames is small. Otherwise, the snake could get stuck to other
contours or edges near the optimal. Some points in the snake are still prone to errors, like get-
ting trapped into the shadow of the pedestrian. This is solved applying restrictions to the re-
gion of search or assigning more stiffness to the model.
Initial Final
position snake
Fig. 3. [Top] Detail of a frame. In gray, the active contour from the previous frame is de-
picted. In white, the extracted contour. [Bottom] From (a) to (f) a complete sequence of
frames is shown.
This algorithm has been tested on images taken by a trinocular camera mounted on a
vehicle. A detail of the on-board vision system is shown in (Fig. 1). Some of the ob-
tained results are exposed in (Fig. 3).
Acknowledgements
References
1 Information Society Technologies for Transport and Mobility. Achievements and Ongoing
Projects from the Fifth Framework Programme: Office for Official Publications of the
European Communities, 2003. <http://europa.eu.int>.
3 Gavrila D.M.; Kunert M.; Lages U. “A multi-sensor approach for the protection of vulner-
able traffic participants-the PROTECTOR project”. IEEE Instrumentation and Measure-
ment Technology Conference. 2001.
5 Kass, M.; Witkin, A.; Terzopoulos D. “Snakes: Active Contour Models”. Int. J. Comp.
Vision, 1(4), pp. 321-331, 1988.
6 Williams, D.J.; Shah, M. “A Fast Algorithm for Active Contours and Curvature Estima-
tion”. CVGIP: Image Understanding, 55(1), pp. 14-26, 1992.
Fast Road Sign Detection using Hough Transform for
Assisted Driving of Road Vehicles
1 Introduction
Traffic sign detection and recognition has experimented increasing research interest in
the last times, this is due to the importance of improve safety in road vehicles. Drivers
sometimes miss signs because of distractions or lack of concentration. Our system
must alert the driver using voiced warning. There are four types of traffic signs in the
traffic code: prohibition, warning, obligation and informative. The most important
traffic signs are prohibition ones; therefore they have priority to be detected in this
work.
To detect a traffic sign in an image, the algorithm follows these steps:
1. Obtaining edge image using Canny method [1].
2. Looking for contours into the edge image.
3. Contours-filtering using aspect constraints.
4. Applying Hough transform [2] to filtered contours.
2 Contours Information
The method used for edge detection is Canny method; this method has been extremely
influential in many applications. Numerous implementations of edge-detectors based
on Canny’s idea have been developed. Canny described a method for generating an
edge-detector using an optimization approach and showed how to use the technique to
generate a robust-edges-detector. Canny’s method preserves contours, what is very
important for detecting traffic sign using shape information, because they are closed
contours. The contours are filtered using aspect constraints. In this sense, circular
contour has similar width and height, prohibition traffic signs are circular, and
therefore their contours will fulfill the aspect constraints. The Hough Transform is
applied to filtered contours, thus the computation-time is reduced. The search area of
traffic signs is focused in the right side of the image as it is showed in Fig. 1.
Fig. 1 Sequence of real road image, filtered edge image into the area of search and Canny
image.
3 Hough Transform
The classical Hough algorithm [3] can be easily extended to find any curves in an
image that can be expressed analytically in the form f(x, p) = 0 [4]. Here, x is a point
in the domain of the image and p is a parameter vector. We use the Hough algorithm
for circle location. A circle in the xy-plane with center (X, Y) and radius r can be
expressed as:
f(x, p) = (x - X)2 + (y - Y)2 - r2 = 0 (1)
Where the parameter space, p = (X, Y, r), must be quantized. The accumulator matrix
a is the representation of the quantized parameter space. For circle detection the
accumulator a will be a three-dimensional matrix with all entries initially set to 0. The
entry a(Xr, Ys, rt) is incremented by 1 for every feature point (xi, yi ), in the image-
domain, contained in the circumference with centre (Xr, Ys) and radius rt, with a
precision margin by ε, used to compensate quantization error when digitize the image
[5]. This condition is expresed:
|(Xr – xi)2 + (Ys – yi)2 – rt2| < ε (2)
In order to reduce the computation time we use contour information, so, every feature
point (xi, yi ) is searched inside a restricted area in every filtered closed-contour.
3 Results and Conclusion
Fig. 2 Sequence of real road images under adverse weather conditions where speed limit is
detected.
An algorithm for real-time detection of traffic signs is carried out based on Hough
Transform. The algorithm has been empirically tested under adverse weather
conditions. We detect the road sign when its closed contour increases the accumulator
matrix a over one threshold, this condition is easy to fulfill for circular shapes. If
several circular shapes are detected it will not be a problem since a neural network
will make a fine classification and validation of the detected road signs. As a future
work this method for circular sings detection can be easly extended to searching
triangular traffic sings using the Hough Transform for lines expressed as:
x cos(β) + y sin(β) = r (3)
A triangular traffic sing is delimitated by three straight lines. A straight line is
expressed analytically in the form f(x, p) = 0, where p(r, β) is the parameter vector.
The accumulator matrix a will be a two-dimensional matrix. The rest of this method is
very similar to circular traffic-signs detection. Our system works between 5 and 50
frames/s, depending of the amount of shapes detected.
References
P. Kopacek
Extended Abstract
The field of robotics is one of the most innovative in the last decade. Conventional
industrial robots from the late 70´s are now only a tool on the production level. One of
the oldest dreams of the robotic community – intelligent, mobile, cooperative as well
as humanoid robots – starts to become reality not only because of the rapid
development of “external” sensors and micro- and nanotechnology.
External sensors (e.g. visual, auditive, force-torque…) combined with micro
drives, embedded systems,… offer intelligent robots the possibility to see, hear,
speak, feel, smell like humans. Compared with conventional, unintelligent, industrial
robots, intelligent robots fulfill new, innovative tasks in new application areas.
There are three “starting” points for the development of intelligent robots:
Conventional, stationary industrial robots; mobile, unintelligent platforms (robots)
and walking machines.
Stationary industrial robots equipped with external sensors are used today for
assembly and disassembly operations, fluelling cars, cleaning of buildings and
airplanes, humanitarian demining ... and have been the first “intelligent” robots.
Mobile platforms with external sensors are available since some years and cover a
broad application field. The core of each robot is an intelligent mobile platform with
an on-board PC. On this platform, various devices, like arms, grippers, transportation
equipment, etc., can be attached. Communication between the „onboard PC“ and the
„supervisory PC“ is carried out by radio-based networks - communication with the
environment can be accomplished by voice, beep or bell.
The application of “agents” in production automation is closely related to
“holons”. Following the results of TC5 of the worldwide ims initiative a holon is an
identifiable part of a system that has a unique identity, yet is made of subordinate
parts and in turn is part of a larger whole. This similar to the usual „fuzzy“ definition
of an agent. The features of a „technical“ agent are:
• Ability to optimize one or more processes simultaneously
• Autonomy – an agent can made own decisions to stabilize to current state and to
improve this state to reach to optimum of the whole system.
• Communication and interaction – for reaching this optimum each agent must
interact with the other agents in the MAS .
The ultimate target to be reached would be a robot that possesses capability
approaching that of human beings - autonomous robot agents. Leaving such an ideal
robot as a goal for the future, intermediate robots that only satisfy a limited selection
of the most requisite functions should still find good use in human society. Among the
faculties cited above, mobility is the most indispensable feature for a service robot.
Walking machines or mechanisms are well known since some decades. Usually
they have 4 to 6 legs (multiped) and only in some cases 2 legs (biped) – walking on
two legs is from the view of control engineering a very complex (nonlinear) stability
problem. Biped walking machines equipped with external sensors are the basis for
“humanoid” robots.
It was an old dream to have a personal robot looking like a human. Main features
of a humanoid robot are
• biped walking
• voice communication – speech recognition
• facial communication
The main feature of a real human is the two legged movement and the two legged
way of walking. Much research has been conducted on biped walking robots because
of their greater potential mobility. On the other side they are relatively unstable and
difficult to control in terms of posture and motion.
Currently there are worldwide two categories of two legged humanoid robots
available:
“Professional” humanoid robots developed by large companies with a huge
amount of research capacities. Examples are: the Honda robots ( P1, P2, P3, ASIMO)
– with the idea to assist humans in everyday working, the SONY robots ( SDRX –
3,4,5) – with the background to serve mostly for entertainment, leisure and hobby or
in the future as personal robots.
“Research” humanoid robots: There a lot of such robots currently available or in
the development stage e.g. approximately worldwide more than 500 University
institutes and research centres are active in this field. The robots of this category a
usually prototypes developed by computer scientists to implement methods of AI,
image processing, theoretical scientists from mechanics implementing and testing
walking mechanisms, control scientists to implement new control strategies, social
scientists to implement human machine interfaces (HMI) for an efficient
communication between humans and humanoid robots.
In addition these intelligent robots – especially mobile platforms and humanoid
robots - are able to work together on a common task in a cooperative way . The goal
are so called “Multi Agent Systems – MAS”. MAS consist of a distinct number of
robots (agents), equipped with different arms, lifts, tools, gripping devices, ... and a
host computer. A MAS has to carry out a whole task e.g. assemble a car. The host
computer divides the whole task in a number of different subtasks (e.g. assembling of
wheels, windows, brakes,...) as long as all this subtasks can be carried out by at least
one agent. The agents will fulfill their subtasks in a cooperative way until the whole
task is solved.
In industry intelligent robots will work together with humans in a cooperative way
on a common working place.
One of the newest application areas of service robots is the field of entertainment,
leisure and hobby because people have more and more free time. In addition modern
information technologies lead to loneliness of the humans (teleworking, telebanking,
teleshopping,....). Therefore service robots will become a real “partner” of humans in
the nearest future. One dream of the scientists is the “personal” robot. In 5, 10 or 15
years everybody should have at least one of such a robot. Because the term personal
robot is derived from personal computer the prices should be equal.
In the paper the present state will be discussed, selected application described and
an outlook on future developments will be given.
Current and Future Trends and Challenges in Robot
Soccer
Extended Abstract
Robot soccer has evolved into a very dynamic and competitive field within the last
few years. It provides excellent research and benchmarking opportunities in a diverse
variety of fields, e.g. multi-agent systems, robot control, sensor fusion, intelligent
control, communication, image processing, mechatronics and artificial intelligence.
Many different robot soccer leagues now exist. The scope ranges from truly tiny,
centrally controlled robots in FIRA NaroSot (4.5 x 4.5 x 5 cm) to humanoid, fully
autonomous robots in humanoid leagues (up to 180 cm in height). At the same time,
the game play of the different leagues can be everywhere between highly autonomous
and research-oriented to high-speed and entertainment-oriented.
The robot soccer league most strongly dedicated to entertainment and edutainment
is currently FIRA MiroSot. This paper will present the current state of robot soccer
within FIRA MiroSot and upcoming trends and challenges.
In the FIRA MiroSot league, the size of a robot is limited to 7.5 x 7.5 x 7.5 cm, which
is very small considering the tasks it has to fulfill. The league started in 1996 with 3
vs 3 robots playing on a field sized 150 x 130 cm. To get closer to “real” soccer, the
number of players on the field slowly grew (5 vs 5, 7 vs 7) along with the field. It has
now reached 11 vs 11 robots on a 440 x 220 cm field.
During the early days of the league, the robots’ mechanical designs, their control
and the complex demand of the strategy were limiting factors. Since then, it has
evolved into a very dynamic, high-speed game with robots reaching speeds of up to
3.5 m/s during game play. Alongside with now refined strategies, this provides a very
entertaining experience to spectators.
Due to the very small size of the robots, they must be supported by a host computer
(usually an off-the-shelf PC) which receives a picture of the field from a camera
mounted about 2.5 m above the field. The host is responsible for image processing
and strategic decisions. It transmits – via a radio link– movement information to the
robots on the field, which they execute, thereby closing the control cycle.
While some problems robot soccer posed have been solved in the last few years
[1], many still exist. At the same time, constant change in the league adds new prob-
lems – especially the change to 11 vs 11 robots has provided many new challenges to
the teams, which will be presented in the following.
The 11-vs-11-field is sized 440 x 220 cm. The cameras are currently mounted at 2.5
m above the field, which means that strong fish eye lenses are necessary to get an im-
age of the entire field. That strong fish eye effect is very undesirable, since it impedes
exact position calculations of the robots. Raising the camera to a higher level is not
possible, since this would rule out too many locations for playing. That means that
most teams playing the 11-vs-11-league must revert to dual camera approaches,
where every camera records one half of the field [2]. Insights into the challenges and
solutions to this problem will be presented.
All FIRA MiroSot robots are controlled by the host computer via a radio link. Cur-
rently, most teams use BiM series radio modules manufactured by the English com-
pany Radiometrix [3]. These transmit in the 433 or 869 MHz ISM bands with speci-
fied gross data rates of up to 160 kbps. Unfortunately, due to various reasons, the net
data rate will be around 35 kbps in this application, with the modules being halfdu-
plex only.
With 22 robots on the field playing robot soccer, the limits of the Radiometrix
modules are now reached. New solutions like differently designed radio hardware
transmitting in different frequency bands as well as updated and more stable protocols
are needed. Selected alternatives to the current communication links will be pre-
sented.
Originally, only 6 robots (3 vs 3) were playing during a FIRA MiroSot game. This de-
sign put emphasis on control and vision problems, because the matter of strategy was
deliberately kept simple. One goalkeeper and two robots actually playing the game
did not allow for any refined strategic moves. This has strongly changed during the
last few years. Strategy design will become much more difficult within FIRA MiroSot
because of the recent extension to 22 (11 vs 11) robots on the field [4]. Simple ap-
proaches as outlined above will not suffice and much more complex strategies will be
needed that get much closer to “human” soccer. Complex team based behaviours like
passes that require a highly sophisticated strategy and coordination will be of advan-
tage to the team mastering them.
Some insight into formerly and currently possible strategy design(s) will be pre-
sented.
References
1. Kim, J.-H., Kim, D.-H., Kim, Y.-J., Seow, K.-T.: Springer Tracts in Advanced Robotics,
Vol. 11: Soccer Robotics. Springer Verlag, Berlin Heidelberg New York (2004)
2. Weiss, N., Jesse, N.: Towards Local Vision in Centralized Robot Soccer Leagues: A Robust
And Flexible Vision System Also Allowing Varying Degrees Of Robot Autonomy. Proc.
FIRA World Congress 2004, Busan, Korea (2004)
3. 869/914MHz High Speed FM Radio Transceiver Module BiM3 Data Sheet. Radiometrix
Ltd, Watford, UK (2004)
4. Hildebrand, L., Michalski, C., Valentin, H., Wickrath, M.: Strategy Implementation For
Mobile Robots Using The Pipes & Filters Architecture. Proc. FIRA World Congress 2003,
Vienna, Austria. (2003)
Strategy and communication in robotic soccer game
Abstract. We describe the key components of our game strategy for robot
soccer [3] in detail. The game can be represented as a trajectory in so-called the
virtual grid. The virtual grid generally allows us to reduce data volume for easy
description of player motion and subsequently for controlling the game or for
learning game strategies. The natural coordinate system is provided by accurate
optical sensing of the subject position using lens for optical transformations and
the CCD camera. This natural coordinate system can be easily mapped to a
virtual grid.
Using the virtual grid, it is possible to describe the position and direction of the robot
using an alphanumeric description see [1,2]. The primary virtual grid can be divided
to (2, 4, 8, . . .) parts, which creates secondary virtual grid.
The tercial grid — strategy grid — depends on the partition of the game field (the
left-right wing, the central field, and transversely the attack-defense field and the
central field). In the discrete frame samples it is possible to study movements and
movement strategies of the robots.
Fig. 1. Alphanumeric representation of robot position and movement using virtual grid.
The game strategy can be dynamically changed based on the game progress (i.e. the
history and the current position of the players and the ball). The game progress can
be divided in time into the following three ground playing classes (GPC):
• GPC of game opening (GPCO)
• GPC of movements in game site (GPCS)
• GPC of game end (GPCE).
The game progress, especially in the GPCS class, can be also divided into the
following two game playing situations (GPS):
• GPS of attack (GPSA). The interactions of simple behaviors cause the
robots to fall into a V-formation where the ball is in motion roughly towards
the opponent’s goal.
• GPS of defense (GPSD). When the ball is not moving roughly towards the
opponent’s goal, the robots move around it to form an effective barrier and
to be in a good position for recovery.
Each GPC has its own movement rules. The classes GPCO and GPCE consist of
finite number of possible movements that are determined by initial positions of
players and the ball. The class GPCS has virtually unlimited number of possible
movements. The movements are determined by the current game situation (GPS) and
by the appropriate global game strategy (in next GGS). The movement of the
particular robot is determined by the current game class and situation, and also by the
robot role. For example, the goalkeeper’s task is to prevent the opponent to score a
goal. His movements are in most cases limited along the goalmouth near of goal line.
The preferred movements are in goal line direction. The preference of these
movements comes from the particular GGS, where the goalkeeper prevents to score a
goal in the way of moving in the position between the central goal point and the ball
(or the expected ball position). The preference of other movement directions is
created using GPSA, where the movements of goalkeeper secure kicking the ball
from the defense zone.
Each strategy is stored in one file and currently consists of about 50 basic rules.
Furthermore the file contains following metadata:
• Information about the name of strategy
• the algorithm to strategy choosing
• the author responsible for current strategy
• the date of last modification
• the size of strategic grid
• strategic rules
Each rule consists of five records:
• The rule ID and description (e.g. Rule 1 "desc"),
• the coordinates of our players in strategic grid (e.g. .Mine a5 a5 a5 a5 a5),
• the coordinates of opponent's players in strategic or virtual grid
(e.g. Opponent k5 j2 i5 h7 i8),
• the ball coordinates in virtual or strategic grid (e.g. .Ball j8)
• strategic or virtual grid positions of the move (.Move a5 i4 g5 i6 j8).
In current algorithm, the .Mine coordinates are not important for movement rule
selection, but they can be used in the future. We believe that the file system for
strategies is an advantage. From observation of opponent's strategy a new set of rules
can be written, without necessity of program code modification. Furthermore, there
is a possibility of automatic strategy (movement) extraction from running game.
.Strategy test 1
.Author "Vaclav Snasel"
.Date "1.5.2004"
.Size 10 8
.Rule 1 "Attack1"
.Mine a4 b3 c1 d1 d4
.opponent a2 b3 c2 d3 d4
.Ball a1
.Move a4 b3 c1 d1 d4
.Rule 2 "Attack2"
.Mine a4 b3 c11 d1 c10
.Opponent a2 b3 c2 d3 a1
.Ball a1
.Move a4 b3 c1 d1 a11
There exist two main criteria in the Strategy selection process. The selection depends
on opponent’s coordinates and ball position. The strategy file contains rules,
describing three possible formations suggesting danger of current game situation.
The opponent’s team could be in offensive, neutral or defensive formations.
Furthermore, we need to weigh up the ball position risk. Generally, opponent is not
dangerous if the ball is near his goal. The chosen rule has minimal strategic grid
distance from current configuration of players and ball.
Optimal movements of our robots are calculated by applying minimal distance from
strategic grid position. The goalkeeper and attacking player, whose distance is
closest to the ball are excluded from strategic movement and their new position is
calculated in exact coordinates
Acknowledgements
The Czech Academy of Sciences partially supported this project within the grant
“Strategic control of the systems with multiagents“.
References
1. Bernatik, R., Horak, B., Kovar, P.: Quick image recognize algorithms. In: Proceeding
International workshop Robot-Multi-Agent-Systems R-MAS 2001. VSB-TU Ostrava
2001, Czech Republic 2001 p.53-58.
2. Horak B., Obitko M., Smid J., Snasel V.: Communication in Robotic Soccer Game.
Communications in Computing 2004: Las Vegas, p. 295-301.
3. Kim, J.; Kim, D.; Kim, Y., Seow, K.: Soccer Robotics (Springer Tracts in Advanced
Robotics), Springer-Verlag, 2004.
Rete algorithm applied to robotic soccer
1 Introduction
2 Foundations
Our proposal is based on the use of Rete algorithm [1] to design a rule-based
expert system. Expert systems design is a branch of artificial intelligence spe-
cialized in development of systems that simulates expert human decision skills.
In particular, rule-based expert systems are composed of a set of facts (knowl-
edge base) and a set of if-then statements (rules). Facts represent information
known (or believed) about the world. And rules describe how that information
could change. The set of rules are fixed during all the execution of an expert
system, but facts can change. Every rule has two parts: the antecedent (a set
of conditions about facts) and the consequent (a set of actions that modify the
knowledge base adding or deleting facts). One of the key features of rule-based
expert systems is the possibility of increasing the knowledge (adding new rules)
without losing the information in the knowledge base.
The expert systems work in an infinite loop. In each cycle they check the
antecedent of every rule. If all the conditions of one of them are true the rule is
activated (and placed on the agenda). The agenda is the list of all rules which
have their conditions satisfied and have not yet been executed. After checking
all the rules, one of the activated ones (that are on the agenda) is selected
to be triggered. The selection method is called conflict resolution strategy, and
depends on the implementation of the system. There are several of them, such as
assigning a priority to every rule, choosing the most recent activated, the most
times triggered, etc.
In every cycle of simulation the system has to check all the new set of facts
and rules that are activated. A direct implementation of an algorithm checking
all the conditions of every rules would be too inefficient.
Rete algorithm takes advantage of two facts [4]. First, most of the facts of
an expert system don’t change from one cycle to another. Thus, it doesn’t check
all the rules every cycle, but it remembers past facts test results, so only new
facts are tested. And second, as several rules can share part of their antecedents
a network is created to minimize the number of tests to be made in each cycle.
It all results in a more efficient algorithm (on the average). The figure 1 shows
an example of optimization of two rules.
?X ?Y ?Z ?X ?Y ?X ?Y ?Z
+ + +
+ +
3 Proposed architecture
The architecture modifies the proposal of Peter Stone in [2]. It, basically, uses
some information common to all the players in a team (called locker-room agree-
ments) and the information received from the world (partial information) to
update the internal state of the agent (Figure 2). With those internal beliefs
a directed acyclic graph (known as external behaviors) selects the action to be
made (an example is shown in figure 3).
Agreements Internal
state
Information Actions
Real world
We propose two modifications of this model. First, all the information re-
ceived from the world is stored as low-level facts. But it’s also generalized as
high-level information relative to objects in the field, in such a way that small
changes in the world won’t lead to changes in the high-level perception of it. Of
course, this generalization must be according to soccer concepts: for example, a
set of positions of players can be generalized in a “offside” fact.
And the second is the use of Rete algorithm to choose the action to be per-
formed. In this case the antecedents of the rules must check high-level properties
(so the algorithm won’t have to be re-testing a lot of rules every cycle), but the
consequent (that chooses the action to be made) can use low-level facts in order
Behavior Face Ball()
...
Behavior Communicate()
...
References
1. Forgy, C. Rete: A Fast Algorithm for the Many Pattern/Many Object Pattern
Match Problem. The Journal of Artificial Intelligence, Vol. 19, 1982, pp. 17-37.
2. Stone, P. Layered Learning in Multi-agent Systems. PhD thesis, School of Computer
Science, Carnegie Mellon University, 1998.
3. Robocup team. Soccer Server System. http://sserver.sourceforge.net.
4. Giarratano & Riley. Expert systems: principles and programming, third edition.
PWS Publishing Company, 1998.
Vision system design and accompanying problems in
robot soccer
Extended Abstract
There are many ways of detecting moving objects using color cameras. However, the
vision systems based on color information proved to be more simple, robust and faster
than most of other recognition methods as stated in [2], [8] and [11]. Sargent et al. [8]
developed a fast real-time vision system with the aid of a special hardware accelerated
system, which only makes sense when software optimizations or accelerations are not
possible. A more reliable vision tracking of moving objects can be achieved by using
robust statistics and probability distributions. A good example of the latter is given in
the color-based face tracking implemented by Bradski [1]. Bruce et al. [2] suggested a
fast vision system for mobile robots by means of efficient color segmentation and a
two-pass connected region determination algorithm. Another important contribution
to the robot soccer vision design was introduced by Wyeth et al. [11], with special
consideration given to the robustness of varying playground illumination conditions.
This paper presents a vision system for mobile robots tracking with the stress given
to the accompanying problems in vision system design such as noise filtering, camera
lens distortion and non-uniform illumination. The basic operation of a global vision
system based on color information is introduced. It is used for mobile robots control
in well-defined and fast dynamic robot soccer game. Therefore simple and efficient
algorithms should be used to optimize the processing time needed for robots
estimation.
Further on an approach to improve an already existing vision system performance
under bad operating conditions is presented. Some fundamentals and solutions to
accompanying problems in vision system design for mobile robot tracking are
presented. Besides methods for filtering and improvement of identified noisy data the
two main factors which deteriorate the performance are dealt with, namely, non-
uniform illumination and camera lens distortion. For the former the problem area and
its origins are focused on and a solution for its compensation by applying
multiplicative component defined by illumination plain is given. The latter consists of
two steps. In the first, radial lens distortion fundamentals are discussed. The suggested
solution for its verification is realized by a geometry model of lens projection. The
second step covers the perspective distortion originating from the tilt of the camera.
For its correction an efficient and robust method of vanishing point detection is
applied. Both correction methods contribute to a vision system performance if
implemented in the appropriate manner.
Applicability of the presented approaches is confirmed on a robot soccer test bed.
Robot soccer is a fast dynamic game and therefore needs an efficient and robust
vision system. To improve the results of the robot soccer vision system the proposed
camera calibration and non-uniform illumination correction algorithm are
implemented. The lens correction method successfully corrects the distortion caused
by the camera lens, thus achieving a more accurate and precise estimation of the
object position. The illumination compensation improves robustness to irregular and
non-uniform illumination which is usually present in real conditions.
References
1. Bradski, G.R.: Computer Vision Face Tracking For Use in a Perceptual User Interface.
Intel Technology Journal. 2nd quarter (1998) 1-15.
2. Bruce, J., Balch, T., Veloso, M.: Fast and inexpensive color image segmentation for
interactive robots. In: Proceedings of IROS-2000. Takamatsu Japan (2000) 2061-2066.
3. Fangi, G., Gagliardini, G., Malinverni, E.S.: Photointerpretation and small scale
stereoplotting with digitally rectified photographs with geometrical constraints. In: CIPA
congres 2001. (2001) http://cipa.icomos.org/papers/2001-.htm.
4. Likar, B., Maintz, B.A., Viergever, M.A., Pernuš, F.: Retrospective shading correction
based on entropy minimization. Journal of Microscopy. 197 (2000) 285-295.
5. Pajdla, T., Werner, T., Hlavač, V.: Correcting Radial Lens Distortion without Knowleadge
of 3-D Structure. Technical report TR97-138. Faculty of Electrical Engineering, Czech
Technical University, Praha, Czech Republic (1997).
6. Perš, J.: Kovačič, S. Nonparametric, Model-Based Radial Lens Distortion Correction Using
Titled Camera Assumption. In: Wildenauer, H., Kropatsch, W. (Eds.): Proceedings of the
Computer Vision Winter Workshop 2002. Bad Aussee Austria (2002) 286-295.
7. Rother, C.: A new Approach for Vanishing Point Detection in Architectural Environments.
In: Proceedings of the 11th British Machine Vision Conference (BMVC'00). Bristol UK
(2000) 382-391.
8. Sargent, R., Bailey, B., Witty, C., Wright, A.: The importance of fast vision in winning the
First Micro-Robot World Cup Soccer Tournament. Robotics and Autonomous Systems. 21
(1997) 139-147.
9. Tsai, R.Y.: A versatile camera calibration technique for high-accuracy 3D machine vision
metrology using off-the-shelf TV cameras and lenses. IEEE Journal of Robotics and
Automation. 3(4) (1987) 323-344.
10. Tuytelaars, T., Proesmans, M., Gool, L.V.: The cascaded Hough transform. In: Proceedings
of International Conference on Image Processing (ICIP ‘97), Vol. 2. Washington (1997)
736-739.
11. Wyeth, G.F., Brown, B.: Robust Adaptive Vision for Robot Soccer. In: Billingsley, J. (ed.):
Mechatronics and Machine Vision in Practice. Research Studies Press UK (2000) 41-48.
12. Jetto, L., Longhi, S., Venturini, G.: Development and Experimental Validation of an
Adaptive Extended Kalman Filter for the Localization of Mobile Robots. IEEE
Transactions on Robotics and Automation. 15(2) (1999) 219-229.
Time-optimal collision avoidance of
automatically guided vehicles
Albert L. Schoute
Extended abstract
v1 = v2
progress of time
d2
trailing delay
v1 vehicle
Collision
states bumper
α
sticking
v2 blocking
d1 ∆
v2
v2
v1
v1
α
β
“meeting state” “greeting state”
References
1. de Groot, R.M.: Dynamic traffic control of free-navigated AGV’s. Master thesis, Univer-
sity of Twente (1997).
2. Hopcraft, J., Schwarz, T., Sharir, M.: On the complexity of motion planning for multiple
independent objects; PSPACE-hardness of the warehouse man’s problem, Journal of Ro-
botic Research 3(4) (1984) 76-88
3. Vloemans, V.P.M.: Autonomous mobile robot collision avoidance through central plan-
ning. Master thesis, University of Twente (2003).
4. Bennewitz, M., Burgard, W., Thrun, S.: Finding and Optimizing Solvable Priority Schemes
for Decoupled Path Planning Techniques for Teams of Mobile Robots, Robotics and
Autonomous Systems, Vol. 41 (2002)
5. Schoute, A.L., Bouwens, P.: Deadlock free traffic control with geometrical critical sections.
Proceedings CSN94, CWI Amsterdam (1994) 260-270 (see
http://wwwhome.cs.utwente.nl/~schoute)
6. MI20 robot soccer website, http://hmi.cs.utwente.nl/robotsoccer
! "#$%'&)(*+#(**("
,.- +("*/ - 0 - 132#(*.465"7+#8
vxwzyR{)|#}{~}'^yR@
y
?@:!OM:QH>Fr^:!OPTH>?@FQOP=^HJ;>FIBLH>FOM=FIH>OG:HKACBWIFJ?A<>xQOM;J;>:OP=@D@:=BQ:x7:<
Hd::=AM=:RFH>F=@Y FIOPWOMYPFQAMWORD@:WOPTJOCQlAMWIW:DHJ;>FrA;JH>FIH>:GR=AJ:!F= BFOPWIOPYPFr
QAMWB;KACF=@"AM=@D¡AQOC;J;>:OP=@D@F=@Y¢AMLH>OP=@OP*OPL@"L@WIH>FAMYP:=^H"H>:¤£VE¦¥3§@¨F=¡AM=
L@=@:<RBWOM;>:D©:=^Fr;>OP=@:=^Hª ?@F3QOM;J;>:OP=@D@:=@Q:6QAM=«:T¬OM;>AMWFIX:D AMABF :QH>Fr^:
AC@FI=@Y6T®;>OPH>?@:BFOPWIOPYPFIQlAMW7H>: §^!H>OA*ACH>?@:6AH>FQAMWnO¯DB:W7OPT)H>?B:QOM;J;>:<
OP=@D@F=@YxL@WIH>FAMYP:=^HH>:°*E¦¥3§ª ?@:D@Fr;>:Q<H>FOP= §^H>O6EG¥3§F±H>?@:ACH>?
:6ACH>FIXlAH>FOP=6OPTH>?@:±BFOPWIOPYPFQAMWORD@:WAC=BD6H>?@:±;>:<^:<;>J:D@FI;>:QH>FIOP=xEG¥3§H>O §^6QAM=
:#F=^H>:;J@;>:<H>:D!AMH>?@:#D@:;>FrPACH>FIOP=!OPT~A3FIOMH>:Q1?@=@FQAMWORD@:WOPTnH>?@:#L@=BD@:<;>WIFI=@Y*EG¥3§ª
?@:*WOMYPFQAMW²A;JH3OPT
H>?B:*EG¥3§³O¯DB:W+FACJ:D«OP=«WOPYMFQAMW+´@:<;>F=@YPzH>:3OMT
D@FIHJ;>FIBLH>:DxWOMYPFQTµOM;EG¥3§ª OF*OM;JHKAM=^H=@OMH>FIOP=@¶RF=^H>:=^H>FOP=AC=BD6;>:e:QH>FIOP=7¶^±FWIW
:#JL@:QHOPT7H>?B:#D@:<^:WOMB:=^HzOPT)AJL@FIHKACBW:6AH>?B:6ACH>FIQlACWT¬OM;>AMWFIXlACH>FIOP=7ª OH>?@FI
:=@D:QAM=;>:JOM;JH²H>O:<FIH>F=@Y#OM;J·OP=WOPYMFQAMWR´@7:<;>F=@YP+T¬OC;
QOROM:<;KACH>F=@Y#;>OC7OCH>AM=@D
ACYP:=^H>)±?@FQK?D@:lAMWI)±FrH>?3H>?@:U=@OMH>FIOP=OPTAMQ:<AM=@DH>FI:UD@:<7:=@D@:=^H~WOPYPFIQlACW^T¬OM;>L@WAMª
?@FIA;>FJ:+F=±?BACH+FI+QAMWIW:DYP:=B:<;>FQ\ORD@:WF=@Y@;>F=@QFIBW:\T¬OM;+AQOM=BHJ;>L@Q<H>FI^:\D@:JQ;>Fr@
H>FIOP=©OMT\JQ:=BAC;>FOP3OPT\OMBFIW:6AMLH>OP=@OPOPL@AMYP:=^H>x£¬;>OMOMH>1¨±±?B:<;>:*H>?@:WIORQlACWWOPYPFIQlACW
HKAH>:3ACQ:3OPT+AM=³AMYP:=^HF±ORD@:W:D"PA;>F®ACBWIF=D@::=@DB:=@Q:3OMT²H>?@:3QOM;J;>:7OM=BD@FI=BY
*Ol^::=^H>ª¸\OP=@Q:<;>=@F=@Y¹¯L@:H>FIOP=@zOPT²ACL@H>OM6ACH>:D;>:AMJOP=@F=@YAM=@D"D@:DBL@Q<H>FOP=F=!H>?B:
WIOPYPFIQlAMWHKACH>:#ACQ:OPT)EG¥3§6AM=@DQO¯OM:;KACH>FI=@Y;>OMOMH>\:QAM=AC@BWI!*:H>?@ORD@T®;>OP
ROPWFIQ\QOPLHKACH>FIOP=AM=@DQOP*BLH>:;+AMWIYP:<B;KAzF=3H>?@:\T®;KAM:<OM;J·3OPTWOPYPFIQlACW^´@:;>FI=BYM
AC=BD 6AM=^RiMAMWILB:D WOPYPFIQPª+º#L;=@:»YP:=@:<;KAMWUO¯DB:W
T¬OM;EG¥3§¦FIAMJ:D OP=$QACH>:YPOM;J
H>?@:OM;J*AM=@D6YP:=@:;KACWBQlAH>:YPOC;>FQAMW@J:6AM=^H>FQªº#TTµL;JH>?B:<;
AMJFIQ±FI=¯H>:<;>:H\FI=6EG¥3§FIH>?B:
FIJJL@:OMTUW:AC;>=@F=@Y±FrH>?¦;>::Q<HH>O"AC;JH>FI´BQF®AMW²=B:L;KAMW+=B:<HdOM;J·¦AC@BWFIQlAH>FOP=@H>O"AMYM:=^H
RH>:AC=BD¦;>OMOMH>FQª7¼ :*LBJ:A"=@:<»YM:OP*:HJ;>FIQP¶7QlAH>:YPOC;>FQAMWORD@:WF=@YAC@@;>O^AMQK?
±?@:<;>:W:AC;>=@F=@YQlAC=6:#TµOM;>6AMWIWIF=^H>:<;J@;>:H>:DxF=6H>:;>*\OPT7OM;JB?@FJªR¥#QQOM;>D@F=@Y3H>O
H>?@:QR:<;>=@:H>FIQB;>FI=@QFrWI:*OPTUT¬:AMJFrFIWFrHd^¶7FrHFFI=¦H>?B:Q:=¯H>:<;OPTUOPL;OC;J·³H>O"AMQ1?@FI:^:
FI*BW::=¯HKAH>FOP=@OPT)AMLH>OP=@OP*OPL@;>OMOMH>\AMJ:D!OP=xH>?@:B;>OC7OMJ:DxBFOPFI:<H>FQEG¥3§
*ORD@:WVª
¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ M ª
! " # $ %
¢{ >{n@{)|#
{~
& ('
)*,+-*/. jp`feaKbdbeo021 *(3 oul]rcdfebda>fdfe]40(1 *(5 o1h6^jl[ *87:9 a*febd]<;¯ob fe]gfea*c hP[¯o ;^cea>=Vulpr]Io!]r[¦cdhM[¯o ;Pfe]<?
febeo[^Ac @]rcecd]rjl[ *(BDC kjlbE6Z\[^]`_Ca>bdce]gfihF1)bdaKcdcG0 BDC kjlbE6H0I JJI *
I *LKM*DKOt^N [Cf 9 aKb *P hPsRaKbd[^a>fe]Q? B [Cfejlprjluh o[R6 7 beo[^c4S t^[R?>fe]rjl[¯op B ;RaKbeofe]rjl[^c *UTWV4X Y<XAZ V4[]\ Y
^ X _a`bcedgfDhH\jilkm2bnigY<kporqnX Yskut vxwzy{fGiA\}|\}o~AYuYsk 0j;^t^s^p * ]r[mMaKpguk =¬jlbduCo[^]rq>]r[^umChPc feaz@c )G I0
m ;¯ob fo
[ ²jMj CGc 0
oc 9 ]g[^ulfejlH[ 0R * Pa* 0R;R; *)G = I0 )l I *
* */3 ]gfdfea>beolt^a>b *P prj ?]lza 6;a>Eb ?Kza ;Pfe]rjl[GcdhMcdfeza @c *X bf>|\}YWXO~A|ncedgYuY V<Zjdg|Rcrn}]cd>_M 0 )) 0
\j * R I } I 0I JJ )*
R* *(3 ]gfdfeaKbeot^aKb *7:9 a*prjlcecjlkUcdaKprsk =VsRjlt^R[ 6¯olbd]gaKc felj olEb 6^co![^a>t^bdj @jlprza ?Kt^prolb#f 9 aKjlb h³jlk
Ac ? 9 ]rqKj ; 9 bdaK[^]ro *TWV4X R Acdg_M 0I I J } I )l 0I JJ *
* *}3 ]gfdfeaKbeot^aKb * @#s¯opIoR[ ?Ka~jlkPupg]IoQp =V[^aKt^bdjl[¯op]r[CfeaKbe o ?>fe]gj[z]r[±cdhM[¯o ;^cdaKc : o ;jlcecd]rs^pra @za ? 9 =
o[^]rAc @'jlk^f 9 a ;¯of 9 j ; 9 hPcd]rjlprjluh±jlk¯s^Q] ;jlpIob 6^]rcdjEb 6^a>b *LdgbfEX g[gVdg|nc4Vu]c 0 ) J <)Gl IJ 0 I JJ R*
*rR* 1)k®olpgqKulbeolk *B [Go6ulaK[^aKbeop+[^jfe]rjl[³jlkUo 9 t^prp * 8 [ bcX _ \ cdE¡m2fE\[zc4V4[]\ Y(¢WdA\ >X |nVu|Z}o
Rk mg\ Y £EZfE\¤\}|¡¦¥-kA§¨\ |Z©wdE¡ zk kDªdA« c-\ |¡O¬©X |XAZ fE\>`nj-Vu|©R _OiAX}YpV4[ ^ X}_a`bce\ c4VX}|Ro
`nf>V®|jZd>fxqd>f>Y<\zZ§FVdg|RoLdg¯±°X f]² 0 )GR*
)k®oprqKulbeok [ ulaKj 3a>febd] ol[ 3fej Rjlprjlul] 1opbda1ocejl[^]r[^u]r[bdjlsRjlfe] Kc
*rR* 1 *{B @ Q? R6 ; <? Q? * |R|\ Y rXD¬©\ c®j³
dg_ \}cV4[zL\}|¡ f>cV ´2[gV\}Y~A|ncedgYuYpVQZjdg|[]d 0 )G I })Gµ 0 )G *
µ*rR* 1 )k®oprqKulbeok a >j[ >a Pfjk
prjlu] 1op ¯sRaKbd]r[^ulco[ ¯sRaKbda ¦pgjul] Kolp Kjl[febdjpgpraKbdc [
*©7:9 F? R? z; Q? 2¶ R6¶ z6 <? 2? *
m2fAXG[]d]dA¡ Vu|Z ^ (°2· ¸¹¹¹lo bZb]caºG³E»}¸ o¸¹¹¹Go(hV}A¼d Zjd>oT,dgY ZVub_Ok _Od>f>V[A\ |x~]|R]cVucbced
XDm2j}]V4[zEo ~m ^ X |GzdgfAd>|[>dam2fEXG[>d]dA¡}V®|jZAorqnX Y<ku½ºl¾w¸¹¹»]lo``Hkut v¾³Et ¿¾kÀ¥-kp¬©kp¥rbniAX Vu
w4Á:¡jk 0IJJJ *
*rR* 1)k®oprqKulbeok j ^aKpr]r[^u >j[^[^a >fe]rjl[^]rcdf"[^a>f +jlb Mc 1ofeaKuljlbd] 1op zulaKj a>febd] «oc a Jfec
*Â3 j6 Ã? g? E ? <? 0 @ <? E; z?
fej ob ^c j 3j jb ] pga1obd[^]r[^u [
Ä l E6
Å 9 @ @ E; 9 Q? RÅ Æ *à m2fEXG[>d]dA¡}V®|jZ ^ (°2· ¸¹¹ ¾}o bZ bAcO»»}³E»tlo
¸¹¹}¾ oÀh{V}dE¼ Zd]oT,dgY ZVub_Ok _Odgf>V4[]\}|O~]|R]cVuc4bcedrXmj}]V4[zAo ~m ^ X |Gzd>f]d>|[>dm2fEXG[>d]dA¡ Vu|jZAo
qnX Y<kº}» vw¸¹¹gÇ}kÀ¥-kp¬©kp¥rbniAX Vu w4Á:¡jk 0IJJ R*
) J *rR* 1 )k®oprqKulbeok [ jul] Kolp 7]rsRaKbd]r[^ulc*o[ tPfej 3ofea a ^t >fe]gj[$]g[ lo [h µ_loprt^a
*ÈB ÊÉ <? aË R6 + @ z6Ì z6 R? 3 = g6
jlul] Kc*Z\cd]r[^u b lj s^[^aKb +oceaKc
É <? K N x *Í¢WdgÎlVu]c\¢WdA\ Y []\¡jd>_MV4\Ì¡d ^ Vdg|[zV4\ Ao¤dgf>Vd ¡d
¬©\ cedg_à Ï\ c4V4[]\ Ao/¢ ^ ¬ 0 µ 0I JJ R*
))*r3*
"jMjlp ^bd] ^ula [ [Cfebdj ^t >fe]rjl[fej t^pgfe]rolulaK[fUmChPc fea c j [ $]rpga>h Mm j[^c Bf
<6 <6 *{+ j6 R? 3 z@ *{ 9 ¦
ÑÐ :É A6H0
I JJjI *
!"#%$&
'!
()$+*+,,&-/.
012+3 4"#
57682 $+*8 9:
"#;$&
<(
?= >A@CBED'FG/HJILKNM'O1PQI1H
RTSQUAVWYX[ZSQ\]X/^`_bac^]Z;Ued?X[SQW/f1gQhiSQ\egQSjAk%\ehml]SQWon[hmXqp2^`_bf?VrisQtedeWou
vxw<y{z}|#~z}~8?y
E&y
BED%FD'IL@>{ED'K>1HIP>1>1&ILPTDILPDT@>1FED'PFO>1DF2FEDI]N@T@CBI1FEC
ED@ PD'4A¡4I2@'>1PPD'¢)>1FEFO2P>1>1ILBI£NFEO2BED@'>?F¤O?¥PQILN>1F
¢¦I1@'D;I1T
E>?I1NF
ILFBED§T>1P¢¦I1@'DI1%@>A>1I1NF)¨A©oF4>LBD P%T>1PE'ªABED§P>1>1TIL
I1O?FE
>D'£eD P¡«@'>?F¤O?¥PQILN>?F>?H%BD¬¢>1NFe¦ILPQI1D D'P-I>?Fe2>?H/BDT>1P¢¦I1@'D>?H/BD
P>L)>LxILP¨%®eQI1FE¦I]P¯EP>1KND'ILPNFOILPD
BEDEPD'@'FD'IL@P>L¦KD
ª&BD
Fe£eD'PD2{NFEDILN@'EP>1KND'+I1FE°BED2NFEO?¥EK±I]Po¡P>L¦KD
¨)²³@'K±I1@`I1KcD BE>{!>
D'¢QILKNB´BEDµP>1>1IL¶N
BEDILEEP>eI1@QB´A¡¸·D'F¦I'£¹»ºILPD'FeD'POE¨©YKNDI1E
>
I¼@'>?KND DKi¡½¢¡->?KN@«D'@ PEN>?F¸>?HBDµEPD'@'
NFED'IL@
>{EDK>?H-I1F¶ILP
ILF½H>1P4BD!¦I1H>LP4BED°PD`I]DFe
>1H;BD°FA£?D'PD°NFED'IL@xEP>1KND'¨©oF
>LPD P>¾PD'EPD'DFe!ILF¶D'FePD°P>1>1!ILP¿@'K±I1
>?H¦I1@«NFeD'PD'¢>¾¤¦FE¶I
@'>?KD'D'K¡¸¢¡)>1KNN@@'KN>?D'´H>1PÀ>?KN¥N>1F¶>?H-BED«NFe£eD PDFD'IL@
EP>1KND'¨
Á#FEOI<oT>/¬¢>?FecP>1>1ILP
ªL¨ ÂT¥@QBeD'PO?D PED'>?FE¢PQILD'-BED/EPF@'KD%>#I1@CBED'£eD
BEQI1¢½§B¸BEDBED'K¶>?HI¯@'>?¥D P°ILKNO?D EPQI¯¢¡¢D'ILEK¡FO¯BENµÃP]>1Ä FED P
¦I1D'D B>Ab¨cÅcI]D'PTD
ILD
>§FÆNFe£eD'¢NOeI]N>?FE'ªI1>?FEO«>1BED P'ª@>?FE¢P¥E@ NFEO
I>1PDx@'>?KND Ç«D¢2D ÇEI1KD?¨È©Y2¥PFED>?¥2BIL-BEN-QI1¢KD`IL->£eD P¡BILP
D PH>1PI1FE@DP>L¦KD4§BED'FBED°ED'O1PD'D°>?H;H±PDD'E>?>?HIP>L)>LI]PÉNFE@ PD`ILD'¨
²#FNFeD PD¢FEOx>1D P£?IL>?F!N<B¦I]§BEDD-FA£?D¢OeILN>1F§B>TD°IFIL¥PQI1KbI¡
>
@'>?FE¢P¥E@'<D'FE@CBEILP{§H>LP@'>?¥D'PI1KO?D PQI{¨
²Ê@'>?KD'D'K¡!EiË)D PD'FA#I¡>xPD'EPD'DFe#BD-FD'IL@>{EDKc>?H/IP>1>1;I]P8
>!KD`I]PFBEDx@'>1PPD'¢>?FEFEO°P>L)>L2IL¾§iBBEDBD'K¼>1H<I¥EQILK¡@CBE>?D'F¾@'>?F¹
FED'@'>?FEN¢FD oT>1P!¥EFOI>`TD PH¥EKILP¤@'±I1KcFED¥PQI1KcFED oT>1PÆJ²#Ì#ÌTN2¥EK±I]>1P¨
BEILEEP>eI1@QB°ILKN>PD`I1@QBED<>{>?F<KN§BD'FBED;FA¥E-D'P>?H¬¢>?Fe§NFE@ PD`ILD'¨
BED'D§D'Ç{D PND'FE@D'H>1PD'BD¦I1N/>?HBED§ED`I>@>?¦FED§>1BILEEP>eI1@QBED§B
BED2BE>1D->x@'>1D§B!BED2FE@'PDI1NFEO4@'>?KD'Çio¡e¨©oFµIEKN>1IBEDI¢¥ED'Fe
Fx-¡O1P>?¥bª{Â#¨Í=?D PQILD'ªeT>LPeD>?FBEN%NEDI-¥ENFEO2D'KND'@'D'P>L)>LILPÊo¡{D%I1FE
BED@'>?¥EKN«ED>?FE¢PQILDB¦I]i;NT>1PBe§BENKD@'>?FENED'PFEO¥E@QBI4@'>?-NFIL>?FHÎ>1P
¥EDF@>1F@ PD DILEKNN@IL>?FE'¨
©oFµ>?¥PT>1PFO4O1P>?¥IxFED'³@'>?FEFD'@ N>?FEN¢N2¥EKNIL>1PxJÏKD'Ç®{ >A>?KQª§BEN@QBµ§NKK
EP>`£{NEDI#ÐD'ÇiKND>{>?KEHÎ>1P/2¥EK±IL>?F>?HI;KNILPO?D<@'K±I1>?HFD'¥PQI1KEFED'oT>1P2o¡{D'&BIL
BI£?DFED'¥EP>L¦Be¡>?KN>1O?N@I1K)P>{>1'ªEBI1D'DF
D £eD'KN>1D'I1FEN¦KDDFeD'¾T>LP
<F
EP>?OLPDCQ¨/²#F>`§FÑF>£eDK§I]BD'IL@`ILK#I]EP>eIL@CBHÎ>1P4>{EDKNFEO¾FED oT>1P¯¢P¥E@ ¹
¥PD'BIL¥ED'D BE>{E#HÎP>?@`ILD'O?>1P¡BED'>1P¡!I1FEµO?D'>?D P¡EP>`£{NEDIEPD@'ND
O1¥EDKNFED#HÎ>1P/BD§O?D'FED'P@N¦KDDFeQIL>?Fx>?H)FED'oT>1P{%NFBEDFED ¸2¥EK±IL>?F>{>?KJ¨
©Y@`ILF°DB>§FµBED'>1PD'@`ILKNK¡§B«I1FµILPO?¥ED'FA;H±P>?8@ILD'O?>1P¡BED'>1P¡B¦I]#BD
ILBEDI]N@I1KÈ>{ED'KÈBED'K>PD'E¥@'D-@>1KND Ço¡>?HKNDILPFENFEOxNF!@>?FE@ PD'D-2¥EK±IL¹
>?FE-§BBEN-¢¡¢D'¨ BDÐD'ÇiKNDx2¥EK±IL>?F¾>A>?K§NKKDTD'KNK/¥ED'¾HÎ>1P-BD
PDILD'Fe§>?HcFE¥E¢PNI1K}ILE¦KN@ILN>1FEP>1KND''¨
ª ª ª
ª !" ª # ª $
ª % ª & ª ')(* ª '" ª !) ¨
+¯ z ,zb
Ez}|#zÈ
-
.0/1.32 \eu`SQriSn .5476)89;:=<6)>?@9A<BDCE6=8;9FHGI. f1UeWohm\eu]SWQ3j J%DS KML}^`OW N*PRQcSQWorihm\j -TS
U
U1.
V . Q . QdeIg W1t{SQWou`SQW .@X hi\4Vriu`^`WohmYX WeZ;hin[Ig WeSQ[ n ZWohmX[SQWohmdeZÊY_ de\ [W ]ehi_S ^`^`\ n[tAVOW N`SQhmXShi\eSQn<V`rmu]SteW[V`hmOn P
Ig WeSQb \ aTriShicg WLde\eu`n[nop1noX[SQZ;en d 2 \ 2 riu`^`WohmYX WeZ;higCVracWohmX[SQWohi^`\_N^]W<YX WeSf1^`rmlVtehirihÍX»p
^`_ 2 rmu]cS P
teW[VhigfLp?nYX[SQZnT^]_ Xf dAV'X[hi^`\ecn g .@2 S f dAV'X[hi^`\eSQn<ZV'YX WeSQZVX[hmgCV_S h*i)j?j -TSk
l jEUeU . j k h
Pmj U j .
d X \eu`rihiOn W-X[W[V\en[rV'X[hi^`\2hio\ n1Q . QdeIg W1t{SQWou`SQWQ1j p .q hi1\ N1riSr W dÎHS ]en . gj a%TW ^`\ te\eSQsW Q&VnoSQnV1\ ] 2 U P
UerihigQVX[hm^]\enQTj t}Wo^1gQSQDS ]ehi\eu`n^`_LYX WeS uq\LX[SQWo\AV'X[hi^`\AVrLa^`\e_NSQWoSQ\egQwS v jjxL}SCVWon^`3_ a%TW ^]\ te\eSyW Q&Vn[STn vij
-TS
SU j zwuYfea%j 2 denYX[WohV? j ^^`1\ ]e^`| \ {VYX WeSZ-V'X[higCVrEf1^1gQhmSX»}p ^SQgX[deWo0S J%^X[Sf1SQWohiSQnQ*j ~È^`r .1V
- j
a&VZ#teWo!h ]eu`Sk%\ehml]SQWon[hÍX»p t}WoSn[nCj -"S
S
U jeUeU . j P h . gj -"Sk)l.
j . Q . QdeIg W1t{SQWou`SQW . aTTW ^`\ te\eSQWÈtAVnoSQTn n 2 \-Vriu`^`WohmYX WeZhmgZ;SYX We ^ ]hi\U{^`rmp?\e^`Z;hVrA!h ]eSQV`reYX WeSQ^`WYp .
uq \ > 8;9
)9AC}Bc<3GI9:)<6)>03=GI8RBDC_G1BY:=c@ ? :)GHBIeBY
j&UAVu`SQn -"U h= V j V . R . z/SQ'h ]eSQr
t}deter . a^`Z;U . jeRT^`OW ]eWoSQcg W]OX PRQc^`noX[^`\ Pm^V`\egQV`nYX[SQWQj -TS
U.
h . Q . QdeIg W1t{SQWou`SQW .x2 UeUerihigQVX[hm^]\en^` _ aTTW ^`\ te\eSQWtAVn[SQn/hm\\e^`\ Prihi\eSCVWgQ^`ZUed?X¢V'X[hi^`\AVru]S^]Z P
SX[WYp . 6)81BDCE6)89FI6)>o[Gm1BYFD8;Gr:0FD9;BD<389 xF@:c8`6)YBm1 4r 479;FIBcxBY
* j - h1n )S Uk j -"S
Uk .
. Q . QcdeIg W1t{SQWou`SQWQyj a . ac^]rmrihi\enCj{V1\ ]Q . Zd?X[sQriSQW .s2 riu`SQteW[Vhig§ZSYX We*^ ]en%_N^]W/u`SQ^`ZSX[Wohig§WoSC=V P
n[^`\ehi\eu .xw<3<o4sBD¡)x¢:)C 8R£FH9; 3j jn U -
-TS j -TS
U
U.
¤.0/1. /3. aW[Vhiu .x¥Y<38Y:T
FD8;9:)<8R:47:*¦Y:)89;FDGI.§2 ]1]ehino^]\ P q SQn[rmS| p t}deter . a^ . j -TS
U
¤.
k . z .X Ig N1Z;hiririSQW . a^`\egQSQWo\ehi\eu«YX WeSSZSQWou`hi\eu«Wo^`riS^]_u`SQ^`Z;SX[WYp¼hm\¯\eSQdeWo^`hi\e_N^]WoZV'X[higQn .
uqb \ ¨ 6)O6)>A>©Bc> ¨ Y:TFIBDGIGI9A<*ª«9A<¬@B O6)>`3=GI8BcC_G6)<
¢:)C 8BccGYs®7FD¯=C_9A>A>!BDIs°r6)c8;CE6=<3<1
°r6 GY¯B}BY
)GH ¨ O:TFcBYBY
)9A< ª)GY£®5> GHBD¡T9;BD0F= ¨5 ¦c>©xr:)c8±;°@:)>A>'6)<
I j -TSS
l.
U. z . t . t)V`der .14s:¦Y:)8 6)<39 >'6)8R:=cGI. {uR²¬t}WoSQn[nQjaVZ#teWo!h ]eu`x S {xVn[n[V`Ig W1denoSXoX[nV1\ ]³^^`1\ ]e^`\j
-TS
U*V .
S.0/1. t}_±VrisQu`W[V_ .`´ \ZV'YX WeSQZ-V'X[higQV`rZ*^ ]eSQrihi\euhi\Wo^`t^X[higQn . uq\ wc89 `FH96)>¥Y<38RBD>A> 9!ª*BD<FcB³6)<
3=C}¦Y:)> 9F 6)81BcCE6)8;9FI6=>0¢:=C 8;9A< ª=01;¢6)> C}BD8[1 @;¢6)Co¦IBc>>®7
)GH ¨ O:TFcBIBY
)9A<*ª
GY
µ BYFD8 YB0r:)8BcG[9<¬¢:)C 8Bc@FH9;BD<FcBE¶"·1¶H
c9A< ª*BD|¸BDc>'6Tª j -TSS j .
-Tl.0/1. t}_±VrisQu`W[V_ .´ \2u`SQ^`ZSX[WohmgTV`1\ ]X[^]U{^`ri^`u`higCVrWoSCVn[^`\ehi\euhi\2Wo^`t{^`X[hmgQn .§w<1<6)> G0: 6)8*±
BDCE6=8;9FHG@6=<
wc8;9 xFD9;6=>¥Y<38RBD>A> 9'ª*BD<FIB j -TS n V
k)S j -TU j -TS
Sk .
-
-
.0/1. t}_ÎVrisu]W[V_ . {*^ ]eSQrihi\eu;gQ^`\e\eSQgX[hi^`\ehinYX/\eSX K^`OW N2noX[WodegX[deWoSQTn nAno^]Z;Su`SQ^`Z;SX[Wohig<V`1\ ]gCV'OX P
SQu`^`WohigCVr¦V`noUSgX[n .xw<3<6=> G@: 6)81BDCE6)89FHG@6)<
Ew¥ j j ¤ n V
k)S j l- j V)l
l*V .
-"V .0/1. t}_±VrisQu`W[V_ . {*^ ]eSQrihi\euÆg^]\e\eSgX[hi^`\ehinoX\eSX K^`OW N1Hn n<gCV'X[SQu`^`WohigCVrj<u`SQ^`Z;SX[Wohig«VnoUSQgX[n
d±X["^ K&VOW ]e
n v We^`Z^`Z;^]Wo1U WehmgrmSCVWo\ehi\e1u v g . uqM\ ¨ O:TFcBIBY
=9<*ª
G¹¢y[ºyx» ¼½½)·=@ ª GY8}¾¾=±O¾¿"
¼½½=·) µ 9=OÀB ª BI[BD> ª
9 C}3wC}BDc9FI6=<}¥I<1GI8;9A8 8RB0: ¨ *=GI9FHGY¥ ¨ ¢:)<THBcIBc<FcB ¨ O:TFcBIBY
)9A<*ª
GY
¸3:)>!;¶=¾)Á«m¼½½DÂ=ó à ¦Y:)9AGE®7
* j V
l
l h .
Controller Network for a Humanoid Robot
Edmund S
hierer, Markus Wuerzl
Vienna University of Te
hnology
Institute of Handling Devi
es and Roboti
s
Favoritenstrasse 9 { 11, 1040 Vienna
Tel.: +43 1 58801 31815, FAX: +43 1 58801 31899
email: s
hiererihrt.tuwien.a
.at, wuerzlihrt.tuwien.a
.at
1 Introdu tion
Humanoid robots presently are one of the most ex
iting, most
hallenging and
most attra
tive resear
h targets in the eld of mobile roboti
s. It
overs not only
me
hani
al engineering, but also in
ludes problems from ele
tri
al engineering,
omputer s
ien
e, arti
ial intelligen
e and more often psy
hology.
Me
hani
al engineering solves questions of
onstru
tion, me
hani
al loads
and properties whilst ele
tri
al engineering gives answers to problems like drive
systems and
omponents, power supply and interfa
ing high power devi
es to
pro
essors and
ontrollers. Computer s
ien
e on the other hand des
ribes algo-
rithms for motion
ontrol, feedba
k
ontrol systems, pattern re
ognition, image
pro
essing and so forth. Arti
ial intelligen
e as a spe
ial appli
ation of
om-
puter s
ien
e may be used for high-grade pro
essing of various sensory inputs
and environmental in
uen
es to perform a more sophisti
ated and
onvin
ingly
behaviour.
One of the keyroles in su
h a
omplex system like humanoids is the pro
essing
system for motion
ontrol to guarentee smooth and "realisti
" traje
tories.
The basi
design of a humanoid robot follows the well proven
onstru
tion of
the human body. That is to say it has 2 legs, a torso, 2 arms and a head.
There is a
ertain limit for the number of DOFs a humanoid robot should have
to satisfy fundamental requirements
on
erning motion-sequen
es. A robot like
Pino [1℄ has 26 DOFs, what is in my opinion the minimum for a realisti
nature
of motion (Figure 1).
Ea
h DOF is realized by one motor. Therefore Pino for example has 26 motors
to generate adequate movements of the robot. To keep the overall
osts for
building a humanoid as low as possible standard servo motors
an be used.
They are available at
omparatively low pri
es and fulll most of the items in
Fig. 1. Te
hni
al view of Pino (
ourtesy of Kitano Symbioti
Systems Proje
t) [1℄
the
atalog of requirements for a ele
tri
al drive used in a humanoid robot. One
weak point in su
h a standard hobby servo is the gear box often made from
syntheti
material whi
h should be
hanged to metal toothwheels to improve
life
y
le and torque.
Mobile robots, in parti
ular a humanoid, needs a bun
h of various sensors.
Omitting high-level sensors like visual sensors for obje
t re
ognition and image
pro
essing there is need for
{ equilibrium sensors to balan
e the robot in any arbitrary situation
{ for
e sensors to dete
t
ollisions with obje
ts and obsta
les or
onta
t to
ground
{ angular positioning sensors for determing joint positions
and mu
h more.
A
ontrol system for a humanoid robot shall
onsist of several "levels of im-
portan
e". Like in a human being one level uses the features and pro
essing
apabilities of its underlying level. If the superior level fails by any reason the
lower one
an
ontinue and performs its asso
iated fun
tionality. This kind of
intera
ting levels of
omplexity is known as subsumption ar
hite
ture and guar-
entees operability of the system down to a
ertain limit even if parts of the
ontroller suer a loss. A Low-budget-version of su
h a network
an
onsist of a
master
ontroller, where all the high-grade fun
tions like path planning, image
pro
essing and so forth are performed, and a bun
h of low-
ost
ontroller nodes.
These nodes
an then pro
ess the
ommands given from the master. The bus the
nodes are inter
onne
ted with is the widespread USB [2℄ due to its simpli
itiy
and its availability in modern
ontroller
hips.
5 Con lusion
This paper des
ribes our approa
h to implement a
ontroller network for a hu-
manoid robot like Pino [1℄. It makes it possible to extend step-by-step the pro-
essing
apabilities of a robot only by adding new
ontrollers to the network.
Referen es
1 Introduction
Developing software for controlling robotic systems is costly due to the complex-
ity inherent in these systems. There is a need for tools that permit a reduction
in the programming effort, aiming at the generation of modular and robust ap-
plications, and promoting software reuse. The techniques which are of common
use today in other areas are not adequate to deal with the complexity associ-
ated with these systems [1]. This document presents CoolBOT [2], a component
oriented framework for programming robotic systems, based on a port automata
model [3] that fosters controllability and observability of software components.
CoolBOT [2] is a C++ component-oriented framework for programming
robotic systems that allows designing systems in terms of composition and inte-
gration of software components. Each software component [4] is an independent
execution unit which provides a given functionality, hidden behind an exter-
nal interface specifying clearly which data it needs and which data it produces.
Components, once defined and built, may be instantiated, integrated and used
as many times as needed in other systems.
In CoolBOT, components are modelled as Port Automata [3]. This concept
establishes a clear distinction between the internal functionality of an active en-
tity, an automaton, and its external interface, sets of input and output ports.
Fig. 1 displays the external view of a component where the component itself is
represented by a circle, input ports, ii , by the arrows oriented towards the circle,
and output ports, oi , by arrows oriented outwards. Fig. 2 depicts an example
of the internal view of a component, concretely the automaton that models it,
where circles are states of the automaton, and arrows, transitions between states.
Transitions are triggered by events, ei , caused either by incoming data through
an input port, or by an internal condition, or by a combination of both. Double
circles indicate automaton final states. CoolBOT components interact and inter
communicate each other by means of port connections established among their
input and output ports. Data are transmitted through port connections in dis-
crete units called port packets. Port packets are also classified by their type, and
This work has been supported by the research project PI2003/160 funded by the
Autonomous Government of Canary Islands (Gobierno de Canarias - Consejerı́a de
Educación, Cultura y Deportes), Spain.
usually each input and output port can only accept a specific set of port packet
types.
CoolBOT introduces two kinds of variables as facilities in order to support
the monitoring and control of components: observable variables, that represent
features of components that should be of interest from outside in terms of control,
or just for observability and monitoring purposes; and controllable variables,
which represent aspects of components which can be modified from outside, in
order to be able to control the internal behavior of a component. Additionally,
to guarantee external observation and control, CoolBOT components provide
by default two important ports: the control port and the monitoring port, both
depicted in Fig. 3. The monitoring port: which is a public output port by means
of which component observable variables are published; and the control port,
that is a public input port through which component controllable variables are
modified and updated. Fig. 4 illustrates graphically a typical execution control
loop for a component using these ports where there is another component as
external supervisor.
i1 o1 e 10
c m
e 12
s3 s7 o1
e3 i1
e5 e9
e 11 e 14
s1 e1
s2 e2
s4 e6
s6 e 13
s8 s9
e4 e7
om in ok
in
s5
Fig. 1: e8
Fig. 3: The control
Component port, c, and the
external view Fig. 2: Component internal view
monitoring port, m
Internally all components are modelled using the same default state automa-
ton, the default automaton, shown in Fig. 5, that contains all possible control
paths that a component may follow. In the figure, the transitions that rule the
automaton are labelled to indicate the event that triggers each one, some of
them correspond to internal events: ok, exception, attempt, last attempt and fin-
ish. The other ones indicate default controllable variable changes: ns r , ns re , ns s ,
ns d , np, and nex. Subscripts in nsi indicate which state has been commanded: r
(running state), re (ready state), s (suspended state), and d (dead state). Event
np happens when an external supervisor forces a priority change, and event nex
occurs when it provokes the occurrence of an exeption.
The default automaton is said to be “controllable” because it can be brought
externally in finite time by means of the control port to any of the controllable
states of the automaton, which are: ready, running, suspended and dead. The
rest of states are reachable only internally, and from them, a transition to one
of the controllable states can be forced externally. Having a look to Fig. 5 we
can see how CoolBOT components evolve along their execution time. Basically,
the default automaton organize the life of a component in several phases which
correspond to different states: starting, devised for initial resource allocation;
ready, the component is ready for a task execution; running, here the component
executes its specific task; suspended, execution has been suspended temporally;
end, a task execution has just been completed.
Furthermore, there are two pair of states conceived for handling faulty sit-
uations during execution, which are part of the support CoolBOT provides for
error and exception handling. One of them devised to face errors during resource
allocation (starting error recovery and starting error states), and the other one
dedicated to deal with errors during task execution (error recovery and running
error states).
ns re
ns d
suspended
nc np
external
ns r
supervisor
ns d ns s ns re
control monitoring nc
ok finish
starting ready ns r running end ns d dead
i1 o1 nc exception
component exception np nex
...
...
np
ns r
attempt
ok (∅ | timer)
om ns re
in np ok
starting
error running error
recovery attempt error last recovery
(∅ | timer) attempt
nc
Fig. 4: A typical last
attempt ns d
starting
component control loop error
ns d
2 A Simple Demonstrator
CoolBOT has been conceived to promote integrability, incremental design and
robustness of software developments in robotics. In this section, a simple demon-
strator will be outlined to illustrate how such principles manifest in systems built
using CoolBOT. The first level of this simple demonstrator is shown in Fig. 6
and it is made up of three components: the Pioneer which encapsulates the set
of sensors and effectors provided by an ActivMedia Robotics Pioneer robot; the
PF Fusion that is a potential field fuser; and a Joystick Navigation component
for managing the robot using a joystick. The integration shown in the figure
makes the robot to to avoid obstacles while going to a destination point. The
second and last level of our demostrator is depicted in Fig. 7. Note that the
systems adds two new components, the Sick Laser which controls a Sick laser
range finder and Scan Alignment that performs self-localization using a SLAM
(Simultaneous Localization And Mapping) algorithm [5][6].
Robustness is another important aspect in robotics. Focusing on this aspect
several tests were prepared to show the behavior of the system upon the malfunc-
tuion of any of its components. In the system of Fig. 7, the key components are:
the Pioneer, Sick Laser and the Scan Alignment components. Based on these
three components, two tests were made. The first of them showed the way the
system works whenever any of the components hangs, and the second one is re-
lated to the degradation of the system when the Sick Laser component stopped
running.
PF
FUSION PF
FUSION
sonar positions/
sonars/bumpers high/low priority sonar positions/
odometry commands sonars/bumpers
odometry reset high/low priority
odometry commands
odometry reset
PIONEER
odometry correction
PIONEER SICK
LASER
high/low priority
commands high/low priority
commands scan
odometry correction
JOYSTICK
NAVIGATION
JOYSTICK
NAVIGATION SCAN
ALIGNMENT
map
3 Conclusions
This document describes briefly CoolBOT, a component-oriented C++ program-
ming framework that favors a programming methodology for robotic systems
that fosters software integration, concurrency and parallelism, asynchronous ex-
ecution, asynchronous inter communication and data-flow-driven processing. The
framework also promotes a uniform approach for handling faulty situations.
References
1. Kortenkamp, D., Schultz, A.C.: Integrating robotics research. Autonomous Robots
6 (1999) 243–245
2. Domı́nguez-Brito, A.C.: CoolBOT: a Component-Oriented Programming Frame-
work for Robotics. PhD thesis, Dpto. Informática y Sistemas, Universidad de Las
Palmas de Gran Canaria (2003)
3. Steenstrup, M., Arbib, M.A., Manes, E.G.: Port automata and the algebra of con-
current processes. Journal of Computer and System Sciences 27 (1983) 29–50
4. Szyperski, C.: Component Software: Beyond Object-Oriented Programming.
Addison-Wesley (1999)
5. Lu, F., Milios, E.: Robot pose estimation in unknown environments by matching
2d range scans. Proc. IEEE Comp. Soc. Conf. on Computer Vision and Pattern
Recognition, Seattle, USA (1994)
6. Lu, F., Milios, E.: Globally consistent range scan alignment for environment map-
ping. Autonomous Robots 4 (1997) 333–349
A Homeostatic-Adaptive Approach for Controlling
Robotic Systems ?
Extended Abstract
In this work, we present a hybrid mechanism for controlling a mobile robotic sys-
tem which combines concepts from homeostatic control and adaptive behavior. The
homeostatic control is inspired by an emotional approach consisting of a set of artifi-
cial hormones [2] computed from pre-categorical sensory data and also from high-level
application results. The adaptive behavior is implemented by a fuzzy controller whose
rules dynamically modify several system parameters, using the same structure to con-
trol both low-level hormonal loops and high-level application tasks. The objective of
this proposal is twofold: guarantee an acceptable image quality keeping the perceptual
data into a homeostatic regime, and use the adaptive behavior to obtain a better resource
management and dynamic response.
Figures 4 and 5 show some data extracted from a test performed on an circuit con-
sisting on two straight parallel segments connected by curved sections on both ends.
The collected data correspond to a complete lap around the circuit. Lightning varies
along the path, with a specially dark area near the beginning due to the existence of a
kind of tunnel that the robot must traverse, so that, without homeostasis the task fails.
In Figure 4, the luminance hormone values are represented together with execu-
tion frequency values for luminance and white balance loops. As it is shown, when
luminance hormone separates from homeostatic regime, luminance loop runs faster to
recover image quality as soon as possible, while white balance becomes slower. This
effect is specially relevant when traversing the dark zone, between seconds 75 and 100.
In homeostatic regime, white balance is allowed to execute at a higher frequency while
luminance loop gets relaxed, for example when the robot is on first straight segment
and first curve (seconds 100 to 125).
Figure 5 shows the stress hormone values, its effect on high-level task quality and
the translation on an execution command affecting translational velocity. As we can
see, the curved sections of the circuit provoke an increase in stress level that reflects in
a reduction on robot velocity, for example when approaching and traversing first curve
between seconds 105 and 130. At the middle of dark zone the robot near stops, waiting
for the recovering of image quality by hormonal loops.
References
1. J. Y. Aloimonos. Introduction: Active vision revisited. In J. Y. Aloimonos, editor, Active
Perception. Lawrence Erlbaum Assoc. Pub., New Jersey, 1993.
2. Dolores Cañamero. Modeling motivations and emotions as a basis for intelligent behavior. In
Johnson Lewis, editor, Proceedings of the First Int. Symposium on Autonomous Agents, pages
148–155, New York, 1997. ACM Press.
3. June-Sok Lee, You-Young Jung, Byung-Soo Kim, and Ko Sung-Jea. An advanced video
camera system with robust AF,AE and AWB control. IEEE Transactions on Consumer Elec-
tronics, 47(3):694–699, August 2001.
4. Harsh Nanda and Ross Cutler. Practical calibrations for a real-time digital onmidirectional
camera. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR
2001), 2001.
5. David B. Stewart and Pradeep K. Khosla. Mechanisms for detecting and handling timing
errors. Communications of the ACM, 40(1):87–93, 1997.
A Mathematical Formalism for the Evaluation of
C-space for Redundant Robots
Extended abstract
In simple terms the collision detection problem can be stated as how can one
have a computer determine when two objects are about to collide, or have col-
lided? The first approach is to evaluate this obstacle avoidance in the robot
workspace; this involves the direct kinematic computation for every robot point,
since the kinematic information is required and the movement planning needs to
be done for each link. After the work of Lozano-Pérez [1] the collision detection
can be performed in a simpler way: the planning problem is solved within the
Configuration Space (C-space). The key point in this approach is the fact that
the robot representation is reduced to a single point in this space, independently
of the type of structure [2]. Lozano-Pérez named this planning task findpath [1],
where a path that connects two points without collisions is found in the C-space.
There is a previous task, findspace[1], that must be performed before any path
can be built: obstacles must be represented in C-space, i.e., C-obstacles must be
computed.
Although some manipulators will benefit from the fact that C-space evalua-
tion algorithms will only depend on the dimension of the considered C-space, for
instance, PUMA-like or Stanford robots, it can become an intractable problem
when considering high-dimensional redundant robots.
In this scenario, the advent of new methods that enable simple and faster
evaluations of the C-space is a challenge. In this sense, Kavraki, in [3], proposes
a method, based on the convolution of the obstacles and a mobile robot, to eval-
uate the discrete C-space. In this method the computation time is independent
of the complexity and shape of the robot and the obstacles for a given resolu-
tion of representation. Although it is doubtless a great advance on the way to
simpler C-space evaluations, only 2D mobile robots are considered, thus lacking
in generality.
While a lot of effort has been done dealing with mobile robots, only a few
works [4][5][6][7] are concerned with articulated manipulators. However, the later
kind of robots are a key element if one wishes to find a general method.
In [8] a more ambitious method is proposed by Curto for many types of
structures that include both mobile and articulated robots, where the obstacle
representation in the C-space is established as the convolution product of two
functions that represent the robot and the obstacles, respectively.
Although this is a big step, the method is not general enough, since some
structures —such as redundant robots— can not be considered. A further step
can be taken by an analysis of the proposed general method as applied to artic-
ulated robots, since changing the reference system for each link would provide
a new method that simplifies the computation, thus reducing both the memory
needs and the computational complexity [9].
This abstract presents a new general method for obstacle representation in
the configuration space for redundant robots. The method is based on the an-
alytical deconstruction of the configuration space, i.e., the separated evaluation
of the C-space portion contributed by the collisions of each link in the kine-
matic chain. The systematic application of a simple convolution of two functions
describing each link and the workspace, respectively, is applied. In order to do
that, the transformation of the workspace among reference systems located at
one point of each link is needed; in this step a sound method, well-known in
robotics, is used. As a main advantage, only the free-collision configurations of
the previous links in the chain are examined when calculating the portion of the
C-space for a given link.
Concisely, it is necessary to choose a proper coordinate system to work with,
considering the particular mechanical structure of the robot and relating it with
the particular degree of freedom (fig. 1).
In this way, a robot A [10] is viewed as a set of r rigid objects.The kinematics
of this chain, i.e., the movement restrictions imposed by the joint to each element,
Ai —the degrees of freedom, DOFs—, would determine some regions of the C-
space for these objects.
Taking into account the Superposition Principle, the calculation of CB for
a robot A, a kinematic chain of r links, is done through the union of all the
CBk related to each of the elements of the robot. The computation of every
C-obstacle region must be done through the evaluation of the associated CBk
functions.
with {q1k , · · · , qsk } ⊆ {q1 , · · · , qm }, where {q1 , · · · , qm } are the DOFs associated
to the robot A. That is, being B the function that describes the workspace,
for the k-th element only the subset of configuration variables associated to
it are considered, and, analogously to what is proposed in [8], each of the
CBk (q1k , · · · , qsk ) functions is evaluated as follows
Z
Ak (q1k , · · · , qsk , x1 , · · · , xn )B(x1 , · · · , xn )dx1 · · · dxn (2)
When solving the integral (2), Ak (q1k , · · · , qsk , x1 , · · · , xn ), the function that
represents the robot, is difficult to evaluate, due to its dependency on all of the
DOFs related to itself and to the previous links in the chain. Thus, the purpuse
Fig. 1. Reference systems used in the kinematic chain
R h i
F Ā0 (0,···,0,q(v+1) ,···,qs ) (q(u+1)k ,···,qvk ,ξ(v+1)k ,···,ξn )
k k
(ξ1 ,···,ξv )
k
0
·F [B (ξ(v+1)k ,···,ξn )] dξ(v+1)k ···dξn
(ξ1 ,···,ξv )
k (5)
Subindices (ξ1 , · · · , ξvk ) denote vk -dimensional transform of functions defined
in Rn . Over the rest of coordinates, (ξ(v+1)k , · · · , ξn ), where a convolution can
not be found, the integration is performed.
References
1. Lozano-Pérez, T.: Spatial planning: A configuration space approach. IEEE Trans-
actions on Computers 32 (1983) 108–120
2. Udupa, S.: Collision Detection and avoidance in controlled manipulators. PhD
thesis, Department of Electrical Engineering, California Institute of Technology
(1977)
3. Kavraki, L.E.: Computation of configuration space obstacles using the fast fourier
transform. IEEE Tr. on Robotics and Automation 11 (1995) 408–413
4. Maciejewski, A.A., Fox, J.J.: Path planning and the topology of configuration
space. IEEE Tr. on Robotics and Automation 9 (1993)
5. Newman, W., Branicky, M.: Real-time configuration space transforms for obstacle
avoidance. The International Journal of Robotics Research 6 (1991)
6. Lozano-Pérez, T.: A simple motion-planning algorithm for general robot manipu-
lators. IEEE Journal of Robotics and Automation 3 (1987) 224–238
7. Lozano-Pérez, T., P.O’Donnell: Parallel robot motion planning. In: Proc. of the
IEEE Int. Conf. on Robotics and Automation. (1991) 1000–1007
8. Curto, B., Moreno, V., Blanco, F.J.: A general method for c-space evaluation
and its application to articulated robots. IEEE Transactions on Robotics and
Automation 18 (2002) 24–31
9. Therón, R., Moreno, V., Curto, B., Blanco, F.J.: Assessing methods for the eval-
uation of the configuration space for planar revolute manipulators. Journal of
Computational Methods in Science and Engineering 4 (2004) 149–156
10. Latombe, J.C.: Robot motion planning. Kluwer Academic Publishers, Boston, MA
(1991)
11. Denavit, J., Hartenberg, R.S.: A kinematic notation for lower-pair mechanisms on
matrices. Journal of Applied Mathematics (1955) 215–221
Global Modal Logics for Multiagent Systems:
A Logical Fibering Approach
Johann Edtmayr
Extended Abstract
The problem of combining logics (or logical systems) is a growing area of interest
in modern logic as shown among other things in [4] and [3]. In recent years logic
is used more and more to formalize complex problems in artificial intelligence,
software engineering, and computer science as a whole. In particular the formal
modeling of systems of cooperating agents is of general interest, because in many
studies of complex systems, especially distributed systems, a multiagent model is
used. The formal modeling of the different aspects of systems in this area usually
requires combined systems of logics. A logical system modeling a reactive agent
should be a combined system of logics of knowledge, belief, time, and modal
logics (of action) [3]. In our logical modeling of a multiagent system (MAS) each
agent is represented by its own (local) logic. The global logical model of the
whole MAS is a composition of the local logics of the agents [6].
A very flexible approach for combining logics is the concept of logical fiberings
which was originally introduced by J. Pfalzgraf (cf. [4]). The development of this
approach was influenced on the one hand by the concept of polycontextural
logics, which comes from the philosophical work of Gotthard Günther, on the
other hand by the mathematical theory of fiber bundles. The concept of logical
fiberings provides a framework for the construction of complex logical models as
follows: Different logical systems called fibers, or local subsystems, are arranged
(modeled) over the points of a base space. Mathematically a logical fibering is
defined by a map π : E → B; E is called total space, and B the base space.
The fiber over a point b ∈ B is exactly the preimage set π −1 (b). Each fiber over
the base space can be a single logical system (e.g. a classical first order logic)
or itself a combined system of logics, i.e., a logical fibering. The base space of a
logical fibering can carry an additional structure, for example a communication
network, which represents some kind of interaction or communication between
the local subsystems. J. Pfalzgraf and coworkers have shown among other thing
in [7] and [6] that this flexible modeling language is also suitable to associate a
system of distributed logics to a MAS. In modeling a multiagent system (MAS)
in this way each individual agent obtains its own logic. In this context we speak of
the local logical state space (i.e. local fiber) of an agent. The whole logical model
can be formed by putting together all local fibers thus yielding the corresponding
logical fibering which presents the global logical state space of the MAS.
For the designer of an autonomous artificial agent the relationship between
knowledge and actions of the agent are of great interest. It has become customary
to use epistemic modal logic for the formal study of knowledge and beliefs of
such systems. So far the concept of logical fiberings was not used in conjunction
with modal logics and possible world semantics. A closer study of systems of
distributed modal logics in terms of fibered structures for the formal modeling
of multiagent systems seems to be reasonable (cf. [8]). Thus our research has been
aimed at testing whether it is possible to build a logical fibering, which represents
a global modal logic modeling a multiagent system, consisting of the local modal
logics assigned to the agents. It has turned out that such a logical fibering results
in a formal model of a MAS, which allows to study the knowledge and beliefs
of agents, taking into account the communication between the agents. To show
first results of our research we consider a simple task (manufacturing process)
which is performed by three cooperating robots. A global modal logic, in terms
of a logical fibering of local modal logics, is used for the formal modeling of the
multiagent system consisting of the three cooperating robots. The main result
of this work is that global modal logics can be represented as logical fiberings.
In our future work we intend to study distributed knowledge, common knowl-
edge, and awareness structures within logical fiberings of modal logics (cf. [2]).
A classical deduction problem can be treated symbolically as a corresponding
ideal membership problem using computer algebra support with the method of
Gröbner bases and the Stone isomorphism. This approach was already used for
m-valued logics parallelized by logical fiberings (cf. [5]). Based on existing work
([5], [1]) we intend to analyse if deduction problems in systems of fibered modal
logics can be treated in a similar way.
References
1. J. Chazarain, A. Riscos, J.A. Alonso, and E. Briales. Multi-valued logic and Gröbner
Bases with applications to modal logic. Journal of Symbolic Computation, 11:181–
191, 1991.
2. R. Fagin, J. Y. Halpern, Y Moses, and M. Y. Vardin. Reasoning About Knowledge.
MIT Press, 1995.
3. Dov M. Gabbay. Fibring Logics. Oxford University Press, 1999.
4. J. Pfalzgraf. Logical Fiberings and Polycontextural Systems. 1991. In: Proc. Funda-
mentals of Artificial Intelligence Research, Jorrand Ph., Kelemen J. (Eds.). Lecture
Notes in Computer Science 535 (subseries LNAI).
5. J. Pfalzgraf. On logical fiberings and automated deduction in many-valued logics
using Gröbner Bases. Revista Real Academia de Ciencias, Serie A de Matemáticas,
RACSAM (Royal Academy of Sciences of Spain) Special issue on Symbolic Compu-
tation in Logic and Artificial Intelligence, 98, 2004.
6. J. Pfalzgraf and J. Edtmayr. The concept of logical fiberings: Distributed logics for
multiagent systems. Cybernetics and Systems (EMCSR 04), 1:58–63, 2004.
7. J. Pfalzgraf, V. Sofronie, and K. Stokkermans. On semantics for cooperative agents
scenarios. Cybernetics and Systems (EMCSR 96), 1:201–206, 1996.
8. T. Porter. Geometric aspects of multiagent systems. Electronic Notes in Theoretical
Computer Science 81, 2003.
The H2 Control Problem: Comparison of State-space
and Transfer-function Approaches
Vladimír Kučera 1, 2, 3
1
Czech Technical University in Prague, Faculty of Electrical Engineering
2
Institute of information Theory and Automation, Academy of Sciences of the Czech
Republic
kucera@fel.cvut.cz
Extended Abstract
The H2 control problem consists of stabilizing the control system while minimizing
the H2 norm of its transfer function. Several solutions to this problem are available.
For systems in state space form, and under the standard regularity assumptions,
Doyle, et al. [1] obtained an optimal regulator in observer form by solving two
algebraic Riccati equations. Stoorvogel [6] relaxed the regularity assumptions while
Kučera [3] obtained a solution for systems in descriptor form. For systems described
by transfer functions, Park and Bongiorno [5] employed Wiener-Hopf optimization to
obtain the optimal regulator transfer function via operations with rational matrices.
Meinsma [4] and Kučera [2] derived a solution using operations with proper stable
rational matrices: inner-outer factorizations and proper stable projections.
The aim of this paper is to compare the two approaches. It is well understood that
the inner-outer factorization is equivalent to solving an algebraic Riccati equation.
However, why are the proper stable projections not needed in the state-space
approach?
To fix ideas, the H2 control problem in the “textbook” form is considered. Given a
state-space realization of the plant
⎡ A B1 B 2 ⎤
⎡G G ⎤
G = ⎢ 11 12 ⎥ := ⎢ C1 0 D12 ⎥
⎣G 21 G 22 ⎦ ⎢C D ⎥
⎢⎣ 2 21 0 ⎥⎦
find a proper rational controller R that stabilizes G and minimizes the H2 norm of the
transfer function H from w to z in the standard control system configuration
w z
G
u y
R
3 Supported by the Ministry of Education of the Czech Republic under Research Program
MSM212300013.
It is assumed that (A, B2) is stabilizable, (A, C2) is detectable, the matrix
⎡ A − jωI B 2 ⎤
⎢ C D12 ⎥⎦
⎣ 1
has full column rank for all finite ω, the matrix
⎡ A − jωI B1 ⎤
⎢ C D21 ⎥⎦
⎣ 2
has full row rank for all finite ω, and
T T
∆12 := D12 D12 > 0 , ∆21 := D 21 D 21 > 0.
First G is represented in terms of doubly coprime factorizations over RH∞
G = M −1 N = N M −1 ,
with the denominator factors block triangular
⎡ I M 12 ⎤ ⎡N N ⎤ ⎡N N ⎤ ⎡ I 0 ⎤
M =⎢ ⎥ , N = ⎢ 11 12 ⎥ , N = ⎢ 11 12 ⎥ , M = ⎢ ⎥.
⎣0 M 22 ⎦ ⎣ N 21 N 22 ⎦ ⎣ N 21 N 22 ⎦ ⎣ M 21 M 22 ⎦
Then all controllers RS that stabilize G are parametrized as
RS (W ) = ( X + WN 22 ) −1 (Y + WM 22 )
= (Y + M22 W )( X + N 22 W ) − 1
where X, Y and X , Y are RH∞ matrices that satisfy the Bézout equation
⎡ M 22 − N 22 ⎤ ⎡ X N 22 ⎤ ⎡ I 0⎤
=
⎢ −Y
⎣ X ⎥⎦ ⎢⎣ Y M 22 ⎥⎦ ⎢⎣0 I ⎥⎦
and where W is a parameter that ranges over RH∞.
Finally, two dual projection results are used to minimize the norm, in which the
asterisks denotes conjugation:
(1) Let E and F be matrices with equally many rows, with E in RH2 and F in RH∞.
Suppose that F is inner and F ∗E is in RH2⊥. Then for any RH2 matrix T,
2 2 2
E − FT = E + T .
(2) Let E and F be matrices with equally many columns, with E in RH2 and F in RH∞.
Suppose that F is co-inner and EF ∗ is in RH2⊥. Then for any RH2 matrix T,
2 2 2
E − TF = E + T .
Using doubly coprime factorizations,
H = G11 + G12 RS ( I − G22 RS )G21
= N11 − VN 21
where
V := M 12 ( X + N 22W ) − N12 (Y + M 22W ) .
~ ~
Write N 21 = UN 21 , where N 21 is co-inner and U is co-outer. Let P denote the
~∗
projection of N 11 N 21 on RH2. Then
~
H = H 1 − V1 N 21
~∗
with H 1 = N 11− PN 21 in RH2 and H 1 N 21 in RH2⊥. Therefore,
2 2 2
H = H1 + V1 .
Now
V1 = VU − P = N11K − N12W2
where
N11K := ( M 12 X − N12Y )U − P , W2 = WU .
~ ~
Write N 12 = N 12U , where N12 is inner and U is outer. Denote P the projection of
~∗
N 12 N 11K on RH2. Then
~
V1 = H 2 − N12V
~ ~∗ ⊥
where H 2 = N 11K − N 12 P is in RH2 and N 12 H 2 is in RH2 . Then
2 2 2
V1 = H2 + V .
To summarize,
2 2 2 2
H = H1 + H2 + V
and
⎡ sE − ( A + KC 2 ) B2 + KD22 ⎤ ⎡ sE − ( A + KC 2 ) K ⎤
X =⎢ ⎥ , Y =⎢
⎣ −L I ⎦ ⎣ −L 0 ⎥⎦
⎡ sE − ( A + B2 L) K ⎤ ⎡ sE − ( A + B2 L) K ⎤
X =⎢ ⎥, Y =⎢ .
⎣ − (C 2 + D 22 L ) I ⎦ ⎣ −L 0 ⎥⎦
Take
T −1
K = −( B1 D21 + Q K C 2T ) ∆21
where QK is the largest non-negative definite solution of the Riccati equation
−1
AQ K + Q K A T + B1 B1T = ( B1 D21
T
+ Q K C 2T ) ∆21 T
( B1 D21 + Q K C 2T )T .
−1 −1
Then ∆212 N 21 is co-inner and N11( ∆212 N 21)∗ belongs to RH2⊥. Further take
−1 T T
L = − ∆12 ( B2 QL + D12 C1 )
where QL is the largest non-negative definite solution of the Riccati equation
−1 T
A T QL + Q L A + C1T C1 = ( B2T Q L + D12
T
C1 ) T ∆12 T
( B2 Q L + D12 C1 ).
−1 −1
Then N12 ∆122 is inner and ( N 12 ∆122 )∗ N 11 belongs to RH2⊥. Since N11K can be
−1
obtained from N11 by replacing B1 by K, ( N 12 ∆122 )∗ N 11 K belongs to RH2⊥ as well.
It follows that
1 2 1 1 2
2 2
H = N 11 + N 11K ∆21
2 + ∆12
2 W∆ 2
21
References
1. Doyle, J.C., Glover, K., Khargonekar, P.P., Francis, B.A.: State space solutions to
standard H2 and H∞ control problems. IEEE Trans. Automatic Control 34 (1989)
831-847
2. Kučera, V.: The H2 control problem with internal stability. In: Proc. Joint CCA,
IASIC, and CACSD, Taipei (2004), to appear
3. Kučera, V.: The H2 control problem for descriptor systems. In: Proc. 16th IFAC
World Congress, Prague (2005), submitted
4. Meinsma, G.: On the standard H2 problem. In: Kučera, V., Šebek, M. (eds.):
Robust Control Design, A Proceedings Volume from the 3rd IFAC Symposium,
Prague, Czech Republic. Pergamon, Oxford (2000) 681-686
5. Park, K., Bongiorno, J.J.: A general theory for the Wiener-Hopf design of
multivariable control systems. IEEE Trans. Automatic Control 34 (1989) 619-626
6. Stoorvogel, A.A.: The singular H2 control problem. Automatica 28 (1992) 627-631
Analysis, Synthesis and Mimetization of Nonlinear
Dynamical Control Systems using Mathematica
Abstract.
In this work we report the development of computer-animated mimics of some
nonlinear dynamical control systems frequently used in the literature for
illustration purposes. Mathematically speaking these mimics are discrete-time
nonlinear dynamical control systems, arising from the discretization of
continuous-time nonlinear dynamical systems. Besides considering and
implementing standard linear discretization techniques, we also report some
experiments on nonstandard symbolic discretization of nonlinear dynamical
systems that exploit the advantages of both the Euler polygons method and the
Picard iteration method. The proposed nonstandard discretization techniques
might be described as a triplet (Taylor approximation, Euler sampling, Picard
interpolation), and therefore might be controlled at least threefoldly. In fact, we
go a little further by introducing controls over the sampling time and/or the
number of Picard iterations that depend on a prescribed fitting error criterion.
We used the integrated symbolic, graphic, and numeric capabilities of
Mathematica to develop all our work within a single computation environment.
1 Introduction
That the inner structure of our brain seemed to be better fitted to parallel image
processing than sequential numerical processing is summarized in the universal
saying a single picture worth a thousand words. We may roughly describe the content
of this work as a dynamic upgrade of the previous maxim: a mimic worth a thousand
pictures. More precisely, in this paper we report the development of a set of
computer-animated mimics of physical processes commonly used in control theory as
plants. These mimics allow visualizing both the open-loop dynamics of the plants,
and the performance, advantages, and disadvantages of control laws intended to
modify such open-loop behaviors. Comprehension through visual mimetization has
some obvious advantages over traditional numerical simulation, namely, it is closer to
physical and geometric intuition and, to a great extend, it does not depend on the
order of the systems, since what our brain recognizes is a unique dynamical moving
object. So, visualization allows us to exploit the strong geometric understanding and
integrating capabilities of our brains, instead of relying on their weaker algebraic
manipulation abilities. Contrastingly, visual mimetization might also jeopardize our
fine resolution capabilities concerning discriminating subtle details on the exhibited
dynamics.
Having said that, it is worth to remark that our work is mainly concerned with the
dynamical aspects of controlled dynamical systems, and the mathematics supporting
it. Thus, beyond attractive graphics animation, for the sake of this work a mimic is an
nth-order discrete-time nonlinear dynamical control system (DT-NLCS) [1, 6],
Φ((k+1)T) = F(Φ(kT), u(kT)), Φ(0) = xo; Ψ(kT) = H((Φ(kT), u(kT)) (1)
obtained by discretization of a smooth enough nth-order continuous-time nonlinear
dynamical control system (CT-NLCS) [7]
x’ = f(x, u), x(0) = xo; y = h(x, u) (2)
Obviously, both the discrete-time state-vector orbit Φ and the output-vector orbit Ψ
depend on the discretization method. Besides the Euler, Euler-Picard and Euler-
Taylor-Picard discretization proposals already described in [6], in this paper we
consider modifying the Euler-Picard and Euler-Taylor-Picard discretization methods,
through the introduction of error-based controls over the sampling period and the
number of Picard iterations, to improve the computation time [2].
When the DT-NLCS (1) describes the closed-loop dynamics of a control system,
the trayectories Φ and Ψ also depend on the control laws u, which may have different
linear or nonlinear structures [3, 4, 5]. Anyhow, the associated controllers are linear or
nonlinear dynamical systems, and are themselves also discretized using the proposed
discretization methods.
At its present stage of development, our library of mimics includes a family of
pendulum system: a simple pendulum, an inverted pendulum, an inverted pendulum
mounted on a car, two single inverted pendulums mounted on a car, a double inverted
pendulum mounted on a car, a centrifugal pendulum, a one-dimensional airplane, and
a family of biophysical ionic-channels.
xt
4 xp t
t
To 2 To 3 To 4 To
Figure 1. Graphical comparison between standard Euler polygons discretization method vs the
proposed Euler-Picard discretization method. Curves x(t), xe(t) and xp(t) denote the solution of
the CT-NLCS (2), its Euler polygons approximation and its Euler-Picard approximation,
respectively. The discretized trajectories are obtained by restricting x(t), xe(t) and xp(t) to the
sequence of sampling times 0, To, 2 To, 3 To, … .
References
Extended Abstract
The cost reduction in the hardware implementation of digital control systems
has two main efiects. On the one hand, the fleld of applications is enlarged as
the cost of the control system becomes afiordable with respect to the cost of the
controlled process. On the other hand, the control device is reduced on size lead-
ing to a limitation on the available resources. As a result, the control subsystem
is integrated into the whole controlled system. The control is embedded and the
main features of this kind of systems should be taken into account. The main
purpose of this paper is to review the characteristics of the embedded control
systems (ECS), as well as to outline some results the authors have developed
in difierent related issues. Among the main features, the following ones are the
most relevant from the control viewpoint:
These features require dealing with some issues in designing and implement-
ing the control, were many research groups are working on [12, 13, 11, 1, 16,
15, 22]. For these issues, some previous results of the authors are discussed and
referenced:
And altogether, the integration of control design and computer control im-
plementation is crucial to achieve the best control performances [21]. To consider
the relevance of each control action, and consequently each control task, to mod-
ify accordingly their priorities and distribute the available computation time to
each one, to reduce the delivery time interval for any control action or to balance
the control performance degrading of all the control loops are features already
treated by the authors. Some results can be outlined from the references d com-
puter control implementation is crucial to achieve the best control performances
[4, 14, 5].
Control kernel
In real time embedded systems, an operating system kernel should be provided to
handle the basic requirements of the whole system even working in the emergency
conditions [8]. Similarly, a control solution should be implemented to ensure the
safe behaviour of the system even in the case of resources shortage. This implies
a reduced order model of the plan, a simplifled control algorithm and a reliable
operation. These basic properties should be captured by a kernel representation
of the system in order to apply an essential control. Thus, in this case, the sys-
tem will fall under this control until more resources (options) become available.
Among other emergency conditions, the following circumstances are analysed: i)
The time allocated for control computation is reduced and a safe operation (with
degraded performances) is foreseen. Thus, the sampling period may be enlarged
and only the slow part of the plant model could be considere, and ii) Under an
emergency condition, a quick reaction of the control is expected. The sampling
period is reduced and the slow behaviour of the plant may be considered as
unchanged.
In both cases a reduced model of the plant is desirable and a change in the
model parameters should be implemented. A similar reasoning will be applied if
dealing with the controller model. There are many model reduction techniques
but the real-time transfer between equivalent representations should be inves-
tigated. This is the main theoretical result of this paper. Generally speaking,
the kernel representation implies a reduced order model of the system and a
mechanism to transfer from a normal, extended model to the kernel and vice
versa. For that purpose, a simple algorithm to transfer between models, also
recovering the involved data, should be provided. The transferring approaches
require scheduling support to allow mode changes in the system. A mode change
is initiated whenever the systems detects a change in the environment or in the
internal state that must drive it from one operating mode to another allowing
the use of reduced models and the transfer between models. Several protocols for
mode changes can be found in the literature. A survey on mode change protocols
can be found in [19]. The main requirements to be achieved by the protocols are:
schedulability, periodicity, promptness and consistency.
References
Extended abstract
The purpose of this work is to analyze methods and algorithms to control the
ACUREX solar collectors field of the Plataforma Solar de Almerı́a - PSA). This
solar thermal system is based in parabolic trough technology. Parabolic trough
collectors are solar concentrating devices suitable to work in the 1500 C - 4000 C
temperature range. A parabolic trough collector is made up of a parabolic-
trough-shaped mirror that reflects and concentrates direct solar radiation onto
a receiver tube located in the focal line of the parabola. The concentrated radia-
tion heats a fluid flowing through the receiver tube, thus transforming the solar
radiation into thermal energy in the form of the sensible heat of the fluid. In the
ACUREX solar field the heat transfer fluid used is oil type synthetic which can
be heated up to 300 0 C. The heated oil is stored in an oil tank. This single tank
serves not only as feed tank, transferring oil from the power block to the solar
field, but also as thermal energy storage system, hot and cold oil can be stored
in the same tank because density of oil used change a lot with the temperature
and hot oil remains always in the upper levels of the tank because its density
is much lower than that of cold oil. Traditionally, as the energy source in this
system can not be manipulated, the way of controlling this system is by mean of
manipulation of the oil flow through the solar field in order to achieve a desired
oil temperature level at the outlet of the solar field.
The analysis of methods and algorithms to control the ACUREX solar field
is a starting point to formulate a model for solar collectors field control opti-
mization. The sole criterion of our optimization policy is to maximize the en-
ergy accumulated in the oil storage tank but the optimal strategy (maximum
efficiency of obtaining energy is a major factor) should fulfill some constraints
resulting from physical and technical restrictions. It results in a nonlinear dy-
namic optimization problem with nonlinear constraints, which is considered in
literature as hard.
References
1. ACUREX Field D.A.S. and Control System, Documentation version 2001,Platforma
Solar de Almerı́a, Almerı́a, Spain 2001.
2. Blanco-Muriel M., Alarcón-Padilla D.C., López-Moratalla T., Lara-Coira M.: Com-
puting the solar vector, Solar Energy 70(5):431-441, 2001.
3. Camacho E.F., Berenguel M., Rubio F.R.: Advanced control of solar plants, Ad-
vances in Industrial Control, Springer-Verlag, 1997.
4. Carmona R.: Analisis, modelado y control de un campo de colectores solares dis-
tribuidos con sistema de seguimiento en un eje, PhD thesis, Universidad de Sevilla,
1985.
5. Farkas I., Vajk I.: Experiments with internal model-based controller for Acurex
Field, IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain,
2002.
6. Klempous R,. Maciejewski H., Nikodem M. , Nikodem J., Berenguel M., Valenzuela
. L.,” DATA DRIVEN METHODS AND DATA ANALYSIS OF A DISTRIBUTED
SOLAR COLLECTOR FIELD”, APLIMAT International Confernce Bratyslava,
February, 2005 (to be appear0
7. Solar thermal electricity generation, Lectures from the summer school at the P.S.A.
13-17 July 1998, Almerı́a, Spain.
8. Valenzuela L., Zarza E.: Example of module regulaci.c - software package for the
Acurex field. P.S.A. Doc.No. R06/97 LV, Almerı́a, Spain, 1997.
9. Valenzuela L., Balsa P.: A series and parallel feedfoward control schemes to regulate
the operation of a solar collector field, Serie Ponencias, Proceedings of the 1st Users
Workshop, Almerı́a, Spain, 1997.
10. Vaz F., Oliviera R.: Neves de Silva R.: PID control of a solar plant with gain
interpolation. Serie Ponencias, Proceedings of the 1st Users Workshop, Almerı́a,
Spain, 1997.
11. Zarza E. at all: The DISS Project: Direct Steam Generation in Parabolic Trough
Systems. Operation and Maintenance Experience and Update on Project Status.
Journal of Solar Energy Enginnering Transactions of the ASME, 124(2):126- 133,
2002.
Author Index